before_code stringlengths 30 149k | reviewer_comment stringlengths 16 20.6k | after_code stringlengths 30 163k | diff_context stringlengths 0 22.4k | file_path stringlengths 6 155 | comment_line int64 0 26 | language stringclasses 19
values | quality_score float64 0.07 1 | comment_type stringclasses 9
values | comment_length int64 16 20.6k | before_lines int64 2 4.61k | after_lines int64 2 4.4k | is_negative bool 2
classes | pr_title stringlengths 7 140 | pr_number int64 18 61.2k | repo_name stringclasses 29
values | repo_stars int64 321 56k | repo_language stringclasses 14
values | reviewer_username stringclasses 378
values | author_username stringclasses 739
values | user stringlengths 361 200k | assistant stringlengths 16 20.6k | system stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
# receive tensors
for start, end in zip(reordered_starts, reordered_ends, strict=False):
ret_mask[start:end] = True # set the mask
# receive metadata
metadata_dict = self.broadcast_object_fn(
None, self... | I'm not sure about this. Why are we not allocating from the memory allocator? | if self.metadata.is_first_rank():
# Broadcast total chunk count
chunk_count = len(reordered_chunks)
self.broadcast_object_fn(chunk_count, self.metadata.first_rank)
# Broadcast each chunk's data
for key, memory_obj, start, end in reordered_chunks:
... | @@ -864,6 +850,215 @@ def close(self) -> None:
logger.info("LMCacheEngine closed.")
+ def _process_tokens_internal(
+ self,
+ tokens,
+ mask,
+ reordered_keys,
+ reordered_memory_objs,
+ reordered_starts,
+ reordered_ends,
+ ret_mask,
+ ):
+ ... | lmcache/v1/cache_engine.py | 26 | Python | 0.357 | question | 77 | 51 | 51 | false | [VLLM|CacheEngine]support only save on rank 0 for DS MLA | 1,128 | LMCache/LMCache | 4,864 | Python | YaoJiayi | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [VLLM|CacheEngine]support only save on rank 0 for DS MLA
BEFORE CODE: # receive tensors
for start, end in zip(reordered_starts, reordered_ends, strict=False):
ret_mask[sta... | I'm not sure about this. Why are we not allocating from the memory allocator? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
else:
# NOTE: key should always be in the lookup cache once
# we support it.
location = self.storage_manager.contains(key)
if location is None:
# TODO(Jiayi): Need to refactor P2P as a storage backend to
... | @YaoJiayi This method is abstracted from `retrieve` method without modify any content. | @_lmcache_nvtx_annotate
def health(
self,
) -> int:
"""
Check the health of the cache engine.
return: 0 if healthy, otherwise the error code
"""
return 0 if self.memory_allocator.memcheck() else -1
def close(self) -> None:
"""Close the cache engin... | @@ -874,6 +842,186 @@ def close(self) -> None:
logger.info("LMCacheEngine closed.")
+ def _process_tokens_internal( | lmcache/v1/cache_engine.py | 26 | Python | 0.357 | refactor | 87 | 51 | 51 | false | [VLLM|CacheEngine]support only save on rank 0 for DS MLA | 1,128 | LMCache/LMCache | 4,864 | Python | maobaolong | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [VLLM|CacheEngine]support only save on rank 0 for DS MLA
BEFORE CODE: else:
# NOTE: key should always be in the lookup cache once
# we support it.
location = self.... | @YaoJiayi This method is abstracted from `retrieve` method without modify any content. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
MemoryObjs from the StorageBackends and convert them into GPU KV caches
by GPUConnectors specialized for the serving engine.
It also supports prefetching the KV caches from the StorageBackends.
It relies on the StorageBackends to manage the requests of prefetching
and real retrieval and avoid the c... | Curious why we pass these communication related functions as init arguments. |
class LMCacheEngine:
"""The main class for the cache engine.
When storing the KV caches into the cache engine, it takes GPU KV
caches from the serving engine and convert them into MemoryObjs that
resides in the CPU. The MemoryObjs are then being stored into the
StorageBackends in an asynchronous m... | @@ -74,13 +76,25 @@ def __init__(
memory_allocator: MemoryAllocatorInterface,
token_database: TokenDatabase,
gpu_connector: GPUConnectorInterface,
+ broadcast_fn: Callable[[torch.Tensor, int], None], | lmcache/v1/cache_engine.py | 26 | Python | 0.214 | nitpick | 76 | 51 | 51 | false | [VLLM|CacheEngine]support only save on rank 0 for DS MLA | 1,128 | LMCache/LMCache | 4,864 | Python | YaoJiayi | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [VLLM|CacheEngine]support only save on rank 0 for DS MLA
BEFORE CODE: MemoryObjs from the StorageBackends and convert them into GPU KV caches
by GPUConnectors specialized for the serving engine.
It also support... | Curious why we pass these communication related functions as init arguments. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
)
self.broadcast_fn(tensor, self.metadata.first_rank)
# Create temporary memory object (key not needed for other ranks)
memory_obj = TensorMemoryObj(raw_data=tensor, metadata=metadata)
reordered_chunks.append((None, memory_obj, start, end)... | Raising a performance concern here: are other ranks waiting for this pice transfer? If so, this on-gpu broadcast seems to be redundant. A better approach would be using an on-cpu shared memory broadcast or unifying all local lmcache workers into a single process.
We can postpone handling this though. | reordered_chunks: List of tuples containing [key, memory object, start, end]
ret_mask: Boolean mask indicating which positions have been processed
Side Effects:
- On first rank:
* Broadcasts chunk count and each chunk's combined metadata
* Broadcasts tensor data
... | @@ -879,6 +847,186 @@ def close(self) -> None:
logger.info("LMCacheEngine closed.")
+ def _process_tokens_internal(
+ self,
+ tokens,
+ mask,
+ ret_mask,
+ ) -> List[Tuple[CacheEngineKey, MemoryObj, int, int]]:
+ """Process tokens and populate the reordered lists.
+... | lmcache/v1/cache_engine.py | 26 | Python | 0.643 | performance | 304 | 51 | 51 | false | [VLLM|CacheEngine]support only save on rank 0 for DS MLA | 1,128 | LMCache/LMCache | 4,864 | Python | YaoJiayi | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [VLLM|CacheEngine]support only save on rank 0 for DS MLA
BEFORE CODE: )
self.broadcast_fn(tensor, self.metadata.first_rank)
# Create temporary memory object (key not needed f... | Raising a performance concern here: are other ranks waiting for this pice transfer? If so, this on-gpu broadcast seems to be redundant. A better approach would be using an on-cpu shared memory broadcast or unifying all local lmcache workers into a single process.
We can postpone handling this though. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
Check for external KV cache hit.
Args:
request (Request): the request object.
num_computed_tokens (int): the number of locally
computed tokens for this request
Returns:
the number of tokens that can be loaded from the
external KV ... | As there are two place similar, suggest to abstract a method |
Args:
layer_name (str): the name of the layer.
kv_layer (torch.Tensor): the paged KV buffer of the current
layer in vLLM.
attn_metadata (AttentionMetadata): the attention metadata.
**kwargs: additional arguments for the save operation.
"""... | @@ -777,6 +791,12 @@ def get_num_new_matched_tokens(
token_ids, request.mm_hashes, request.mm_positions
)
+ tags = None
+ if request.sampling_params.extra_args is not None: | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.214 | refactor | 60 | 51 | 51 | false | [CacheEngine] Support separate cache key by user through lmcache tags | 1,200 | LMCache/LMCache | 4,864 | Python | maobaolong | xshwu-ai | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [CacheEngine] Support separate cache key by user through lmcache tags
BEFORE CODE: Check for external KV cache hit.
Args:
request (Request): the request object.
num_... | As there are two place similar, suggest to abstract a method | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
@dataclass(order=True)
class CacheEngineKey:
fmt: str
model_name: str
world_size: int
worker_id: int
chunk_hash: int
tags: OrderedDict = None
def __hash__(self):
return hash(
(
self.fmt,
self.model_name,
self.world_size,
... | Is it necessary to add an addition check `and len(tags) == 0` | torch.float8_e4m3fn: "fp8_e4m3",
torch.float8_e5m2: "fp8_e5m2",
}
STR_DTYPE_TO_TORCH_DTYPE = {v: k for k, v in TORCH_DTYPE_TO_STR_DTYPE.items()}
@dataclass(order=True)
class CacheEngineKey:
fmt: str
model_name: str
world_size: int
worker_id: int
chunk_hash: int
tags: Optional[OrderedD... | @@ -78,10 +79,14 @@ def __hash__(self):
)
def to_string(self):
- return (
+ s = (
f"{self.fmt}@{self.model_name}@{self.world_size}"
f"@{self.worker_id}@{self.chunk_hash}"
)
+ if self.tags is not None: | lmcache/utils.py | 26 | Python | 0.357 | suggestion | 61 | 51 | 51 | false | [CacheEngine] Support separate cache key by user through lmcache tags | 1,200 | LMCache/LMCache | 4,864 | Python | maobaolong | xshwu-ai | Review this pull request change.
Language: Python
File: lmcache/utils.py
PR Title: [CacheEngine] Support separate cache key by user through lmcache tags
BEFORE CODE: @dataclass(order=True)
class CacheEngineKey:
fmt: str
model_name: str
world_size: int
worker_id: int
chunk_hash: int
tags: Ord... | Is it necessary to add an addition check `and len(tags) == 0` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Request id
req_id: str
# Total prompt token length
prompt_len: int
# The token ids that has been scheduled so far
token_ids: list[int]
# The block ids that has been allocated so far
# NOTE: allocated blocks could be more than the number of tokens
# FIXME: need to check whether t... | Optional[OrderDict]? @hickeyma do you know why linter cannot detect this? |
def extract_tags(
config: LMCacheEngineConfig, sampling_params: SamplingParams
) -> OrderedDict:
tags = None
tag_keys = None
if config.extra_config is not None and isinstance(config.extra_config, dict):
tag_keys = config.extra_config.get("tag_keys")
if sampling_params.extra_args is not Non... | @@ -101,6 +112,8 @@ class RequestTracker:
# Multimodal hashes and positions
mm_hashes: Optional[list[str]] = None
mm_positions: Optional[list["PlaceholderRange"]] = None
+ # The request tags
+ tags: OrderedDict = None | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.214 | question | 73 | 51 | 51 | false | [CacheEngine] Support separate cache key by user through lmcache tags | 1,200 | LMCache/LMCache | 4,864 | Python | YaoJiayi | xshwu-ai | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [CacheEngine] Support separate cache key by user through lmcache tags
BEFORE CODE: # Request id
req_id: str
# Total prompt token length
prompt_len: int
# The token ids that has been sc... | Optional[OrderDict]? @hickeyma do you know why linter cannot detect this? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
@dataclass
class SaveSpec:
# Skip already saved tokens
skip_leading_tokens: int
# Whether the scheduler allow us to save the tokens
can_save: bool
@dataclass
class DisaggSpec:
req_id: str
receiver_info: NixlReceiverInfo
is_last_prefill: bool = False
num_transferred_tokens: int = 0
t... | if we have more than one tags, init in here is inflexible |
@dataclass
class SaveSpec:
# Skip already saved tokens
skip_leading_tokens: int
# Whether the scheduler allow us to save the tokens
can_save: bool
@dataclass
class DisaggSpec:
req_id: str
receiver_info: NixlReceiverInfo
is_last_prefill: bool = False
num_transferred_tokens: int = 0
t... | @@ -89,6 +90,16 @@ class DisaggSpec:
tmp_disagg_tracker: dict[str, DisaggSpec] = {}
+def extract_tags(sampling_params: SamplingParams) -> OrderedDict:
+ tags = None
+ if sampling_params.extra_args is not None:
+ if kv_transfer_params := sampling_params.extra_args.get("kv_transfer_params"):
+ ... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.214 | nitpick | 57 | 51 | 51 | false | [CacheEngine] Support separate cache key by user through lmcache tags | 1,200 | LMCache/LMCache | 4,864 | Python | chunxiaozheng | xshwu-ai | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [CacheEngine] Support separate cache key by user through lmcache tags
BEFORE CODE:
@dataclass
class SaveSpec:
# Skip already saved tokens
skip_leading_tokens: int
# Whether the scheduler allow ... | if we have more than one tags, init in here is inflexible | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# TODO(Serapheim): there's a chance we don't want the local
# CPU cache in front of ours. Let's experiment and potentially
# change that in the future.
storage_backends[str(weka_backend)] = weka_backend
if config.gds_path is not None:
gds_backend = GdsBackend(config, loop, me... | `is_normal_backend` sounds a bit weird. why not use `isinstance` instead, such as
```
if isinstance(backend, LocalCPUBackend):
...
else:
...
``` | # TODO(Serapheim): there's a chance we don't want the local
# CPU cache in front of ours. Let's experiment and potentially
# change that in the future.
storage_backends[str(weka_backend)] = weka_backend
if config.gds_path is not None:
gds_backend = GdsBackend(config, loop, me... | @@ -103,4 +103,25 @@ def CreateStorageBackends(
backend_name = str(remote_backend)
storage_backends[backend_name] = remote_backend
- return storage_backends
+ # Only wrap if audit is enabled in config
+ if config.extra_config is not None and config.extra_config.get(
+ "audit_backend_... | lmcache/v1/storage_backend/__init__.py | 26 | Python | 0.786 | suggestion | 157 | 37 | 37 | false | Introduce audit_backend to record audit log of backend | 1,205 | LMCache/LMCache | 4,864 | Python | chunxiaozheng | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/__init__.py
PR Title: Introduce audit_backend to record audit log of backend
BEFORE CODE: # TODO(Serapheim): there's a chance we don't want the local
# CPU cache in front of ours. Let's experiment and potentially
... | `is_normal_backend` sounds a bit weird. why not use `isinstance` instead, such as
```
if isinstance(backend, LocalCPUBackend):
...
else:
...
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
async def move(req: MoveRequest):
try:
event_id = "Move" + str(uuid.uuid4())
msg = MoveMsg(
event_id=event_id,
old_position=req.old_position,
new_position=req.new_position,
tokens=req.tokens,
copy=req.cop... | Why dictionary here? Maybe add a simple comment here? | async def move(req: MoveRequest):
try:
event_id = "Move" + str(uuid.uuid4())
msg = MoveMsg(
event_id=event_id,
old_position=req.old_position,
new_position=req.new_position,
tokens=req.tokens,
copy=req.cop... | @@ -222,19 +222,21 @@ class HealthRequest(BaseModel):
class HealthResponse(BaseModel):
event_id: str
- alive: bool
+ error_codes: dict[int, int] | lmcache/v1/api_server/__main__.py | 26 | Python | 0.214 | question | 53 | 51 | 51 | false | [Controller] Introduce health check by controller | 1,219 | LMCache/LMCache | 4,864 | Python | YaoJiayi | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/api_server/__main__.py
PR Title: [Controller] Introduce health check by controller
BEFORE CODE: async def move(req: MoveRequest):
try:
event_id = "Move" + str(uuid.uuid4())
msg = MoveMsg(
event_id=... | Why dictionary here? Maybe add a simple comment here? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
:param dst_device: the device where the blocking retrieved KV is stored,
could be either "cpu", "cuda", or "cuda:0", "cuda:1", etc.
:raise: RuntimeError if the device is not valid
"""
try:
torch.device(dst_device)
except RuntimeError:
raise
... | How about we introduce a new method `get_listener(self) -> Optional[StorageBackendListener]` so that we don't need to use `hasattr` here.
Can do something like:
```python
if listener := self.get_listener():
listener.on_evict(self, keys)
``` |
:param dst_device: the device where the blocking retrieved KV is stored,
could be either "cpu", "cuda", or "cuda:0", "cuda:1", etc.
:raise: RuntimeError if the device is not valid
"""
try:
torch.device(dst_device)
except RuntimeError:
raise
... | @@ -31,6 +32,20 @@ def __init__(
raise
self.dst_device = dst_device
+ self._listener: Optional[StorageBackendListener] = None
+
+ def set_listener(self, listener: StorageBackendListener):
+ """
+ Set the listener to receive events.
+ """
+ self._listener = l... | lmcache/v1/storage_backend/abstract_backend.py | 26 | Python | 0.857 | suggestion | 251 | 51 | 51 | false | [Bugfix] Fix lookup server registration bug #1189 | 1,237 | LMCache/LMCache | 4,864 | Python | ApostaC | chickeyton | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/abstract_backend.py
PR Title: [Bugfix] Fix lookup server registration bug #1189
BEFORE CODE:
:param dst_device: the device where the blocking retrieved KV is stored,
could be either "cpu", "cuda", or "cuda:0", "c... | How about we introduce a new method `get_listener(self) -> Optional[StorageBackendListener]` so that we don't need to use `hasattr` here.
Can do something like:
```python
if listener := self.get_listener():
listener.on_evict(self, keys)
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
from lmcache.utils import (
CacheEngineKey,
_lmcache_nvtx_annotate,
start_loop_in_thread_with_exceptions,
)
from lmcache.v1.config import LMCacheEngineConfig
from lmcache.v1.lookup_server import LookupServerInterface
from lmcache.v1.memory_management import (
MemoryAllocatorInterface,
MemoryFormat,
... | Let's try not to make `StorageManager` inherit from `StorageBackendListener`.
Alternatively, we can define a new sub class of `StorageBackendListener` (e.g., something like `EvictionListener`) that holds the reference to the storage manager object.
When the `on_evict` function of the `EvictionListener` is called, i... | from lmcache.v1.lookup_server import LookupServerInterface
from lmcache.v1.memory_management import (
MemoryAllocatorInterface,
MemoryFormat,
MemoryObj,
)
from lmcache.v1.storage_backend import CreateStorageBackends
from lmcache.v1.storage_backend.abstract_backend import (
AllocatorBackendInterface,
... | @@ -42,7 +43,7 @@
# TODO: extend this class to implement caching policies and eviction policies
-class StorageManager:
+class StorageManager(StorageBackendListener): | lmcache/v1/storage_backend/storage_manager.py | 26 | Python | 0.643 | suggestion | 492 | 51 | 51 | false | [Bugfix] Fix lookup server registration bug #1189 | 1,237 | LMCache/LMCache | 4,864 | Python | ApostaC | chickeyton | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/storage_manager.py
PR Title: [Bugfix] Fix lookup server registration bug #1189
BEFORE CODE: from lmcache.utils import (
CacheEngineKey,
_lmcache_nvtx_annotate,
start_loop_in_thread_with_exceptions,
)
from lmcache.v1.confi... | Let's try not to make `StorageManager` inherit from `StorageBackendListener`.
Alternatively, we can define a new sub class of `StorageBackendListener` (e.g., something like `EvictionListener`) that holds the reference to the storage manager object.
When the `on_evict` function of the `EvictionListener` is called, i... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def __hash__(self):
if self.tags is None:
return hash(
(
self.fmt,
self.model_name,
self.world_size,
self.worker_id,
self.chunk_hash,
)
)
return... | Hey @sdimitro , I thought you fixed some problems regarding the chunk hash format in #1299. Can you quickly check the code here?
I think we need this modification since the `from_string ` are using something like `int(chunk_hash, 16)`. But just wanted to double check | def __post_init__(self):
tags = None
if self.request_configs is not None:
for k, v in self.request_configs.items():
if k.startswith("lmcache.tag."):
if tags is None:
tags = {}
tags[k[len("lmcache.tag.") :]] =... | @@ -106,7 +106,7 @@ def __hash__(self):
def to_string(self):
s = (
f"{self.fmt}@{self.model_name}@{self.world_size}"
- f"@{self.worker_id}@{self.chunk_hash}"
+ f"@{self.worker_id}@{hex(self.chunk_hash)}" | lmcache/utils.py | 26 | Python | 0.571 | style | 270 | 51 | 51 | false | [Bugfix] Fix lookup server registration bug #1189 | 1,237 | LMCache/LMCache | 4,864 | Python | ApostaC | chickeyton | Review this pull request change.
Language: Python
File: lmcache/utils.py
PR Title: [Bugfix] Fix lookup server registration bug #1189
BEFORE CODE: def __hash__(self):
if self.tags is None:
return hash(
(
self.fmt,
self.model_name,
... | Hey @sdimitro , I thought you fixed some problems regarding the chunk hash format in #1299. Can you quickly check the code here?
I think we need this modification since the `from_string ` are using something like `int(chunk_hash, 16)`. But just wanted to double check | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
self.storage_manager = storage_manager
def on_evict(
self, backend: StorageBackendInterface, keys: List[CacheEngineKey]
):
"""
remove keys from lookup server only if they don't exist in local backends.
:param StorageBackendInterface backend: ... | Rather than defining an internal class, we can have this class defined in `storage_backend_listener.py`. | self, backend: StorageBackendInterface, keys: List[CacheEngineKey]
):
"""
remove keys from lookup server only if they don't exist in local backends.
:param StorageBackendInterface backend: The backend that evicted the keys.
:param List[CacheEngineKey]... | @@ -47,6 +48,34 @@ class StorageManager:
The StorageManager is responsible for managing the storage backends.
"""
+ class _BackendListener(StorageBackendListener):
+ def __init__(self, storage_manager: "StorageManager"):
+ self.storage_manager = storage_manager
+
+ def on_evict(
... | lmcache/v1/storage_backend/storage_manager.py | 26 | Python | 0.5 | suggestion | 104 | 51 | 51 | false | [Bugfix] Fix lookup server registration bug #1189 | 1,237 | LMCache/LMCache | 4,864 | Python | ApostaC | chickeyton | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/storage_manager.py
PR Title: [Bugfix] Fix lookup server registration bug #1189
BEFORE CODE: self.storage_manager = storage_manager
def on_evict(
self, backend: StorageBackendInterface, keys: List[Cach... | Rather than defining an internal class, we can have this class defined in `storage_backend_listener.py`. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
return hash(
(
self.fmt,
self.model_name,
self.world_size,
self.worker_id,
self.chunk_hash,
self.layer_id,
)
)
return hash(
... | @Shaoting-Feng @sammshen Hey folks, can you take a look here to make sure that this change won't impact other functionalities like PD disaggregation? | return hash(
(
self.fmt,
self.model_name,
self.world_size,
self.worker_id,
self.chunk_hash,
self.layer_id,
)
)
return hash(
... | @@ -244,7 +244,7 @@ def __hash__(self):
def to_string(self):
s = (
f"{self.fmt}@{self.model_name}@{self.world_size}"
- f"@{self.worker_id}@{self.chunk_hash}@{self.layer_id}"
+ f"@{self.worker_id}@{hex(self.chunk_hash)}@{self.layer_id}" | lmcache/utils.py | 26 | Python | 0.429 | question | 149 | 51 | 51 | false | [Bugfix] Fix lookup server registration bug #1189 | 1,237 | LMCache/LMCache | 4,864 | Python | ApostaC | chickeyton | Review this pull request change.
Language: Python
File: lmcache/utils.py
PR Title: [Bugfix] Fix lookup server registration bug #1189
BEFORE CODE: return hash(
(
self.fmt,
self.model_name,
self.world_size,
sel... | @Shaoting-Feng @sammshen Hey folks, can you take a look here to make sure that this change won't impact other functionalities like PD disaggregation? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"""
Check if the remote server contains the key
Input:
key: a string
Returns:
True if the cache engine contains the key, False otherwise
"""
raise NotImplementedError
@abc.abstractmethod
async def get(self, key: CacheEngineKey) -> Option... | This seems duplicated with L209 | """
Check if the remote server contains the key
Input:
key: a string
Returns:
True if the cache engine contains the key, False otherwise
"""
raise NotImplementedError
@abc.abstractmethod
def exists_sync(self, key: CacheEngineKey) -> bool... | @@ -130,6 +130,24 @@ async def get(self, key: CacheEngineKey) -> Optional[MemoryObj]:
"""
raise NotImplementedError
+ async def batched_get( | lmcache/v1/storage_backend/connector/base_connector.py | 26 | Python | 0.071 | refactor | 32 | 51 | 51 | false | perf(mooncakestore connector): optimize get/put with zero-copy operations | 1,269 | LMCache/LMCache | 4,864 | Python | maobaolong | xiaguan | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/base_connector.py
PR Title: perf(mooncakestore connector): optimize get/put with zero-copy operations
BEFORE CODE: """
Check if the remote server contains the key
Input:
key: a string
... | This seems duplicated with L209 | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
extra_config.get(
"metadata_server", MooncakeStoreConfig.DEFAULT_METADATA_SERVER
),
),
global_segment_size=file_config.get(
"global_segment_size",
extra_config.get(
"global_segment_size",
... | You have to override this `support_batched_get` method and return true for mooncake | def _unregister_cpu_buffer(self):
"""Unregister CPU buffer."""
if self.registered_buffer_ptr is not None:
result = self.store.unregister_buffer(self.registered_buffer_ptr)
if result == 0:
logger.info(f"Unregistered buffer: {hex(self.registered_buffer_ptr)}")
... | @@ -139,11 +334,277 @@ def __init__(
self.loop = loop
self.local_cpu_backend = local_cpu_backend
+ self.metadata = metadata
+ self.registered_buffer_ptr = None
+
+ # Initialize ReplicateConfig
+ self.replica_config = ReplicateConfig()
+ self.replica_config.replica_... | lmcache/v1/storage_backend/connector/mooncakestore_connector.py | 26 | Python | 0.357 | performance | 83 | 51 | 51 | false | perf(mooncakestore connector): optimize get/put with zero-copy operations | 1,269 | LMCache/LMCache | 4,864 | Python | maobaolong | xiaguan | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/mooncakestore_connector.py
PR Title: perf(mooncakestore connector): optimize get/put with zero-copy operations
BEFORE CODE: extra_config.get(
"metadata_server", MooncakeStoreConfig.DEFAU... | You have to override this `support_batched_get` method and return true for mooncake | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
except ValueError as e:
logger.error("Configuration loading failed: %s", e)
raise
except Exception as exc:
logger.error("An error occurred while loading the configuration: %s", exc)
raise
self.loop = loop
self.local_cpu_backend = local_cp... | The log has been printed in parent class `RemoteConnector`, is the `init_chunk_meta` method need to override here? |
except ValueError as e:
logger.error("Configuration loading failed: %s", e)
raise
except Exception as exc:
logger.error("An error occurred while loading the configuration: %s", exc)
raise
self.loop = loop
self.local_cpu_backend = local_cp... | @@ -152,33 +166,232 @@ def __init__(
self.loop = loop
self.local_cpu_backend = local_cpu_backend
+ self.registered_buffer_ptr = None
+
+ # Initialize ReplicateConfig
+ self.replica_config = ReplicateConfig()
+ self.replica_config.replica_num = 1
+
+ # Set preferred... | lmcache/v1/storage_backend/connector/mooncakestore_connector.py | 26 | Python | 0.5 | question | 114 | 51 | 51 | false | perf(mooncakestore connector): optimize get/put with zero-copy operations | 1,269 | LMCache/LMCache | 4,864 | Python | chunxiaozheng | xiaguan | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/mooncakestore_connector.py
PR Title: perf(mooncakestore connector): optimize get/put with zero-copy operations
BEFORE CODE:
except ValueError as e:
logger.error("Configuration loading failed: %s", e)
... | The log has been printed in parent class `RemoteConnector`, is the `init_chunk_meta` method need to override here? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
n_read,
)
except Exception as exc:
logger.error(f"Reshape failed for key {keys[i]}: {exc}")
memory_objs[i].ref_count_down()
return results
except Exception as exc:
logger.error(f"batch_g... | If we sallow this exception, the vllm will generate wrong output. | logger.error(f"Reshape failed for key {keys[i]}: {exc}")
memory_objs[i].ref_count_down()
return results
except Exception as exc:
logger.error(f"batch_get_into threw exception: {str(exc)}")
# Release any buffers we successfully allocat... | @@ -152,33 +166,232 @@ def __init__(
self.loop = loop
self.local_cpu_backend = local_cpu_backend
+ self.registered_buffer_ptr = None
+
+ # Initialize ReplicateConfig
+ self.replica_config = ReplicateConfig()
+ self.replica_config.replica_num = 1
+
+ # Set preferred... | lmcache/v1/storage_backend/connector/mooncakestore_connector.py | 26 | Python | 0.357 | bug | 65 | 51 | 51 | false | perf(mooncakestore connector): optimize get/put with zero-copy operations | 1,269 | LMCache/LMCache | 4,864 | Python | maobaolong | xiaguan | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/mooncakestore_connector.py
PR Title: perf(mooncakestore connector): optimize get/put with zero-copy operations
BEFORE CODE: n_read,
)
except Exception as exc:
... | If we sallow this exception, the vllm will generate wrong output. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
memory_objs[i].ref_count_down()
return [None] * len(keys)
async def _batch_get_buffer(
self, keys: List[CacheEngineKey]
) -> List[Optional[MemoryObj]]:
"""
Batch get using batch_get_buffer when metadata is stored remotely.
This is used when save_chunk... | This is another place for silent error handling. | async def _batch_get_buffer(
self, keys: List[CacheEngineKey]
) -> List[Optional[MemoryObj]]:
"""
Batch get using batch_get_buffer when metadata is stored remotely.
This is used when save_chunk_meta=True (metadata stored with data).
"""
key_strs = [key.to_string()... | @@ -152,33 +166,232 @@ def __init__(
self.loop = loop
self.local_cpu_backend = local_cpu_backend
+ self.registered_buffer_ptr = None
+
+ # Initialize ReplicateConfig
+ self.replica_config = ReplicateConfig()
+ self.replica_config.replica_num = 1
+
+ # Set preferred... | lmcache/v1/storage_backend/connector/mooncakestore_connector.py | 26 | Python | 0.214 | suggestion | 49 | 51 | 51 | false | perf(mooncakestore connector): optimize get/put with zero-copy operations | 1,269 | LMCache/LMCache | 4,864 | Python | maobaolong | xiaguan | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/mooncakestore_connector.py
PR Title: perf(mooncakestore connector): optimize get/put with zero-copy operations
BEFORE CODE: memory_objs[i].ref_count_down()
return [None] * len(keys)
async d... | This is another place for silent error handling. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Check metadata handling mode to match get behavior
if self.save_chunk_meta:
# Use put_parts with metadata stored remotely
await self._put_with_metadata(key_str, memory_obj)
else:
# Use put_from without metadata (zero-copy)
await self._put_without... | I guess we need to raise the exception so that caller can aware and handling. | await self._put_with_metadata(key_str, memory_obj)
else:
# Use put_from without metadata (zero-copy)
await self._put_without_metadata(key_str, memory_obj)
async def _put_without_metadata(self, key_str: str, memory_obj: MemoryObj):
"""
Zero-copy put using ... | @@ -213,19 +426,69 @@ async def get(self, key: CacheEngineKey) -> Optional[MemoryObj]:
return None
async def put(self, key: CacheEngineKey, memory_obj: MemoryObj):
- # Please use a function like `memory_obj.to_meta()`.
- kv_bytes = memory_obj.byte_array
- kv_shape = memory_obj.g... | lmcache/v1/storage_backend/connector/mooncakestore_connector.py | 26 | Python | 0.357 | suggestion | 77 | 51 | 51 | false | perf(mooncakestore connector): optimize get/put with zero-copy operations | 1,269 | LMCache/LMCache | 4,864 | Python | maobaolong | xiaguan | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/mooncakestore_connector.py
PR Title: perf(mooncakestore connector): optimize get/put with zero-copy operations
BEFORE CODE: # Check metadata handling mode to match get behavior
if self.save_chunk_meta:
... | I guess we need to raise the exception so that caller can aware and handling. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
self.worker_id,
self.chunk_hash,
layer_id,
)
)
return keys
def get_first_layer(self) -> "LayerCacheEngineKey":
"""Return the key for the first layer"""
key = LayerCacheEngineKey(
self.fmt,
... | Why there is a 16 here? @sdimitro | "%".join([f"{k}={v}" for k, v in self.tags.items()]),
)
)
def to_string(self):
s = (
f"{self.fmt}@{self.model_name}@{self.world_size}"
f"@{self.worker_id}@{self.chunk_hash}"
)
if self.tags is not None and len(self.tags) != 0:
... | @@ -124,7 +124,7 @@ def from_string(s):
if len(parts) != 5:
raise ValueError(f"Invalid key string: {s}")
return CacheEngineKey(
- parts[0], parts[1], int(parts[2]), int(parts[3]), int(parts[4])
+ parts[0], parts[1], int(parts[2]), int(parts[3]), int(parts[4], 16) | lmcache/utils.py | 26 | Python | 0.071 | question | 34 | 51 | 51 | false | Misc Upstreaming | 1,299 | LMCache/LMCache | 4,864 | Python | maobaolong | sdimitro | Review this pull request change.
Language: Python
File: lmcache/utils.py
PR Title: Misc Upstreaming
BEFORE CODE: self.worker_id,
self.chunk_hash,
layer_id,
)
)
return keys
def get_first_layer(self) -> "LayerCacheEng... | Why there is a 16 here? @sdimitro | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
super().__init__(dst_device)
self.config = config
self.loop = loop
self.memory_allocator = memory_allocator
self.dst_device = dst_device
assert config.weka_path is not None, (
"Need to specify weka_path for WekaGdsBackend"
)
self.weka_path = ... | Can you explain why you use threadpool instead of `asyncio` ? | super().__init__(dst_device)
self.config = config
self.loop = loop
self.memory_allocator = memory_allocator
self.dst_device = dst_device
assert config.weka_path is not None, (
"Need to specify weka_path for WekaGdsBackend"
)
self.weka_path = ... | @@ -153,14 +153,14 @@ def __init__(
self.put_tasks: set[CacheEngineKey] = set()
self.rand = random.Random(self.dst_device)
+ thread_count = config.gds_io_threads or 4
+ self._thread_pool = ThreadPoolExecutor( | lmcache/v1/storage_backend/weka_gds_backend.py | 26 | Python | 0.571 | question | 61 | 51 | 51 | false | Misc Upstreaming | 1,299 | LMCache/LMCache | 4,864 | Python | maobaolong | sdimitro | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/weka_gds_backend.py
PR Title: Misc Upstreaming
BEFORE CODE: super().__init__(dst_device)
self.config = config
self.loop = loop
self.memory_allocator = memory_allocator
self.dst_device = dst_de... | Can you explain why you use threadpool instead of `asyncio` ? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"nixl_peer_host": {"type": Optional[str], "default": None, "env_converter": str},
"nixl_peer_init_port": {
"type": Optional[list[int]],
"default": None,
"env_converter": _to_int_list,
},
"nixl_peer_alloc_port": {
"type": Optional[list[int]],
"default": None,
... | It's better to make this as an extra config? | "nixl_peer_host": {"type": Optional[str], "default": None, "env_converter": str},
"nixl_peer_init_port": {
"type": Optional[list[int]],
"default": None,
"env_converter": _to_int_list,
},
"nixl_peer_alloc_port": {
"type": Optional[list[int]],
"default": None,
... | @@ -185,6 +185,11 @@ def _to_int_list(
"default": None,
"env_converter": int,
},
+ "gds_io_threads": {
+ "type": Optional[int],
+ "default": None,
+ "env_converter": int,
+ }, | lmcache/v1/config.py | 26 | Python | 0.143 | question | 44 | 51 | 51 | false | Misc Upstreaming | 1,299 | LMCache/LMCache | 4,864 | Python | YaoJiayi | sdimitro | Review this pull request change.
Language: Python
File: lmcache/v1/config.py
PR Title: Misc Upstreaming
BEFORE CODE: "nixl_peer_host": {"type": Optional[str], "default": None, "env_converter": str},
"nixl_peer_init_port": {
"type": Optional[list[int]],
"default": None,
"env_converter... | It's better to make this as an extra config? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
memory_obj.ref_count_down()
return None
return memory_obj
def _load_bytes_from_disk_with_allocation(
self,
key: CacheEngineKey,
path: str,
dtype: torch.dtype,
shape: torch.Size,
) -> Optional[MemoryObj]:
"""
Load byte array... | @sdimitro Sorry, can you remind me how does this memory allocator behave? | )
memory_obj.ref_count_down()
return None
return memory_obj
def _load_bytes_from_disk_with_allocation(
self,
key: CacheEngineKey,
path: str,
dtype: torch.dtype,
shape: torch.Size,
) -> Optional[MemoryObj]:
"""
... | @@ -445,6 +445,81 @@ def _load_bytes_from_disk(
return None
return memory_obj
+ def _load_bytes_from_disk_with_allocation(
+ self,
+ key: CacheEngineKey,
+ path: str,
+ dtype: torch.dtype,
+ shape: torch.Size,
+ ) -> Optional[MemoryObj]:
+ """
+ ... | lmcache/v1/storage_backend/weka_gds_backend.py | 26 | Python | 0.357 | question | 73 | 51 | 51 | false | Misc Upstreaming | 1,299 | LMCache/LMCache | 4,864 | Python | YaoJiayi | sdimitro | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/weka_gds_backend.py
PR Title: Misc Upstreaming
BEFORE CODE: memory_obj.ref_count_down()
return None
return memory_obj
def _load_bytes_from_disk_with_allocation(
self,
key: CacheEng... | @sdimitro Sorry, can you remind me how does this memory allocator behave? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"""
# TODO(Jiayi): Consider adding write-back here.
memory_obj = self.local_cpu_backend.allocate(shape, dtype, fmt)
assert memory_obj is not None, "Memory allocation failed during disk load."
buffer = memory_obj.byte_array
size = len(buffer)
fblock_aligned = si... | this path needs the same try exacept right? | start_time = time.time()
size = len(buffer)
fblock_aligned = size % self.os_disk_bs == 0
if not fblock_aligned and self.use_odirect:
logger.warning(
"Cannot use O_DIRECT for this file, "
"size is not aligned to disk block size."
)
... | @@ -590,8 +590,13 @@ def load_bytes_from_disk(
start_time = time.time()
if not fblock_aligned or not self.use_odirect:
- with open(path, "rb") as f:
- f.readinto(buffer)
+ try:
+ with open(path, "rb") as f:
+ f.readinto(buffer)
+... | lmcache/v1/storage_backend/local_disk_backend.py | 26 | Python | 0.071 | question | 44 | 42 | 34 | false | [Fix] EngineCore shutdown when some .pt files was lost | 1,314 | LMCache/LMCache | 4,864 | Python | sammshen | fourierr | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_disk_backend.py
PR Title: [Fix] EngineCore shutdown when some .pt files was lost
BEFORE CODE: """
# TODO(Jiayi): Consider adding write-back here.
memory_obj = self.local_cpu_backend.allocate(shape, dtyp... | this path needs the same try exacept right? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
},
"extra_config": {
"type": Optional[dict],
"default": None,
"env_converter": lambda x: json.loads(x) if x else None,
},
"save_unfull_chunk": {
"type": bool,
"default": True,
"env_converter": lambda x: x.lower() in ["true", "1"],
},
"blocking_time... | what do you think about rename lmcache_metrics_api_server?
we already have api server for cache controller | "type": Optional[dict],
"default": None,
"env_converter": lambda x: json.loads(x) if x else None,
},
"save_unfull_chunk": {
"type": bool,
"default": True,
"env_converter": lambda x: x.lower() in ["true", "1"],
},
"blocking_timeout_secs": {"type": int, "def... | @@ -217,6 +217,16 @@ def _to_int_list(
"default": "LRU",
"env_converter": str,
},
+ "cache_engine_internal_api_server_enabled": { | lmcache/v1/config.py | 26 | Python | 0.429 | performance | 110 | 51 | 51 | false | Lmcache internal api server for metrics export | 1,318 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/config.py
PR Title: Lmcache internal api server for metrics export
BEFORE CODE: },
"extra_config": {
"type": Optional[dict],
"default": None,
"env_converter": lambda x: json.loads(x) if x else None,
},
"save_u... | what do you think about rename lmcache_metrics_api_server?
we already have api server for cache controller | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Create lookup client using factory
self.lookup_client = LookupClientFactory.create_lookup_client(
vllm_config, config
)
self._unfinished_requests: dict[str, Request] = {}
self._lookup_requests_in_step: list[str] = []
else:
... | Shall we put `vllm_config.speculative_config` here like the first four arguments? |
class LMCacheConnectorV1Impl:
def __init__(
self,
vllm_config: "VllmConfig",
role: KVConnectorRole,
parent: KVConnectorBase_V1,
):
self._parent = parent
self.kv_role = vllm_config.kv_transfer_config.kv_role
config = lmcache_get_config()
self.con... | @@ -520,6 +537,7 @@ def __init__(
vllm_config.cache_config,
vllm_config.scheduler_config,
config,
+ vllm_config, | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.357 | question | 81 | 51 | 51 | false | Support addon num_mtp_layers from num_nextn_predict_layers | 1,320 | LMCache/LMCache | 4,864 | Python | Shaoting-Feng | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Support addon num_mtp_layers from num_nextn_predict_layers
BEFORE CODE: # Create lookup client using factory
self.lookup_client = LookupClientFactory.create_lookup_client(
... | Shall we put `vllm_config.speculative_config` here like the first four arguments? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
VLLM_PARALLEL_CONFIG: Optional[ParallelConfig] = None
VLLM_SCHEDULER_CONFIG: Optional[SchedulerConfig] = None
def _init_lmcache_engine(
lmcache_config: LMCacheEngineConfig,
vllm_config: "VllmConfig",
) -> Optional[LMCacheEngine]:
"""Initialize the LMCache engine by the given model config and parallel
... | sorry, I don't quite understand why are we turning these unto tuples?
for exaple, integration tests are failing right now because of:
```text
(EngineCore_0 pid=143) kv_dtype = get_kv_cache_torch_dtype(cache_config.cache_dtype, model_config.dtype)
--
| (EngineCore_0 pid=143) ... | `LMCACHE_CONFIG_FILE` to load the configuration file. If that environment
variable is not set, this function will return None.
:param lmcache_config: The LMCache configuration.
:type lmcache_config: LMCacheEngineConfig
:param vllm_config: The vLLM configuration.
:type vllm_config: VllmConfig
... | @@ -412,6 +417,11 @@ def init_lmcache_engine(
if LMCacheEngineBuilder.get(ENGINE_NAME) is not None:
return None
+ model_config = (vllm_config.model_config,) | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 1 | bug | 726 | 51 | 51 | false | Support addon num_mtp_layers from num_nextn_predict_layers | 1,320 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Support addon num_mtp_layers from num_nextn_predict_layers
BEFORE CODE: VLLM_PARALLEL_CONFIG: Optional[ParallelConfig] = None
VLLM_SCHEDULER_CONFIG: Optional[SchedulerConfig] = None
def _init_lmcache_engi... | sorry, I don't quite understand why are we turning these unto tuples?
for exaple, integration tests are failing right now because of:
```text
(EngineCore_0 pid=143) kv_dtype = get_kv_cache_torch_dtype(cache_config.cache_dtype, model_config.dtype)
--
| (EngineCore_0 pid=143) ... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
" \033[33mAverage generation throughput (per request): "
f"\033[32m{average_generation_speed_per_request:.4f} "
"tokens/req/s\033[0m\n"
)
print(f" \033[33mAverage TTFT: \033[32m{average_ttft:.4f}s\033[0m\n")
print(f"Time range: {start_time} - {end_time} ({... | The `qps` here is to match the process summary to the various qps values in the run script. It is best to rename the `total_requests / total_time` to something like `real_qps` to avoid overwrite. | " \033[33mOutput tokens per second: "
f"\033[32m{average_generation_speed:.4f} tokens/s\033[0m\n"
)
print(
" \033[33mAverage generation throughput (per request): "
f"\033[32m{average_generation_speed_per_request:.4f} "
"tokens/req/s\033[0m\n... | @@ -521,10 +524,9 @@ def summary(self, start_time: float, end_time: float) -> pd.DataFrame:
pending_queries = len([s for s in self.sessions if s.has_unfinished_request])
start_time = max(self.start_time, start_time)
end_time = min(end_time, df["finish_time"].max())
- qps = self.workloa... | benchmarks/multi-round-qa/multi-round-qa.py | 26 | Python | 0.571 | suggestion | 195 | 51 | 51 | false | [Bugfix][Enhancement] Improve multi-round-qa.py and data_preprocessing.py scripts | 1,321 | LMCache/LMCache | 4,864 | Python | Shaoting-Feng | 1saac-k | Review this pull request change.
Language: Python
File: benchmarks/multi-round-qa/multi-round-qa.py
PR Title: [Bugfix][Enhancement] Improve multi-round-qa.py and data_preprocessing.py scripts
BEFORE CODE: " \033[33mAverage generation throughput (per request): "
f"\033[32m{average_generation... | The `qps` here is to match the process summary to the various qps values in the run script. It is best to rename the `total_requests / total_time` to something like `real_qps` to avoid overwrite. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"""
Allocate a memory object of shape and dtype
evict if necessary. Storage manager should always call
local_cpu_backend.allocate() to get memory objects
regardless of whether local_cpu is True or False
"""
logger.debug(
f"Allocating memory in local cp... | Here's a bit weird. Why does the local CPU backend need to handle `CuFileMemoryAllocator`? This should be for gds backend only.
IIUC, the storage manager should have the correct "allocator backend" to allocate and evict the memory objects. The allocator backend should be gds backend when using gds, which will handle... | self.lmcache_worker.put_msg(
KVEvictMsg(self.instance_id, key.worker_id, key.chunk_hash, str(self))
)
# NOTE (Jiayi): This `return True` might not accurately reflect
# whether the key is removed from the actual memory because
# other backends might still (... | @@ -263,8 +264,10 @@ def allocate(
if memory_obj is not None or not eviction:
return memory_obj
- assert isinstance(self.memory_allocator, MixedMemoryAllocator) or isinstance(
- self.memory_allocator, NixlCPUMemoryAllocator
+ assert (
+ isinstance(self.memory_... | lmcache/v1/storage_backend/local_cpu_backend.py | 26 | Python | 0.786 | suggestion | 381 | 51 | 51 | false | [Bugfix] Add KV Cache format in gds backend | 1,324 | LMCache/LMCache | 4,864 | Python | ApostaC | muma378 | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_cpu_backend.py
PR Title: [Bugfix] Add KV Cache format in gds backend
BEFORE CODE: """
Allocate a memory object of shape and dtype
evict if necessary. Storage manager should always call
local_cpu_... | Here's a bit weird. Why does the local CPU backend need to handle `CuFileMemoryAllocator`? This should be for gds backend only.
IIUC, the storage manager should have the correct "allocator backend" to allocate and evict the memory objects. The allocator backend should be gds backend when using gds, which will handle... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
event_manager=self.event_manager,
lmcache_worker=self.lmcache_worker,
lookup_server=self.lookup_server,
)
# HACK: remove this in the future
# NOTE (Jiayi): This is currently used to support
# dropping the kv cache from the buffer in PD backend
... | if this is hard coded from the metadata, can we move all of this into the GdsBackend and its allocator? it will be able to pack and unpack from there as well |
# HACK: remove this in the future
# NOTE (Jiayi): This is currently used to support
# dropping the kv cache from the buffer in PD backend
# at decoder.
self.remove_after_retrieve = config.enable_pd and config.pd_role == "receiver"
self.use_layerwise = config.use_layerwi... | @@ -152,11 +152,14 @@ def __init__(
self.use_layerwise = config.use_layerwise
self.num_layers = metadata.kv_shape[0]
+ self.fmt = MemoryFormat.KV_2LTD | lmcache/v1/cache_engine.py | 26 | Python | 0.429 | suggestion | 157 | 51 | 51 | false | [Bugfix] Add KV Cache format in gds backend | 1,324 | LMCache/LMCache | 4,864 | Python | sammshen | muma378 | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [Bugfix] Add KV Cache format in gds backend
BEFORE CODE: event_manager=self.event_manager,
lmcache_worker=self.lmcache_worker,
lookup_server=self.lookup_server,
)
# HACK:... | if this is hard coded from the metadata, can we move all of this into the GdsBackend and its allocator? it will be able to pack and unpack from there as well | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"type": Optional[str],
"default": None,
"env_converter": str,
},
"nixl_enable_gc": {
"type": bool,
"default": False,
"env_converter": lambda x: x.lower() in ["true", "1"],
},
"nixl_backends": {
"type": Optional[list[str]],
"default": None,
... | can we make the default UCX like nixl originally does since you are currently passing None into the consturctor if this is unspecified | },
"nixl_role": {"type": Optional[str], "default": None, "env_converter": str},
"nixl_receiver_host": {
"type": Optional[str],
"default": None,
"env_converter": str,
},
"nixl_receiver_port": {
"type": Optional[int],
"default": None,
"env_converter": in... | @@ -158,6 +169,11 @@ def _to_int_list(
"default": False,
"env_converter": lambda x: x.lower() in ["true", "1"],
},
+ "nixl_backends": { | lmcache/v1/config.py | 26 | Python | 0.357 | suggestion | 134 | 51 | 51 | false | [Misc] Allow passing nixl backends when creating nixl_agent in nixl_connector_v3 | 1,343 | LMCache/LMCache | 4,864 | Python | sammshen | jinqinn | Review this pull request change.
Language: Python
File: lmcache/v1/config.py
PR Title: [Misc] Allow passing nixl backends when creating nixl_agent in nixl_connector_v3
BEFORE CODE: "type": Optional[str],
"default": None,
"env_converter": str,
},
"nixl_enable_gc": {
"type": bo... | can we make the default UCX like nixl originally does since you are currently passing None into the consturctor if this is unspecified | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"""
self.nixl_config = nixl_config
self.side_channel = side_channel
if nixl_config.buffer_size > NixlPipe.TRANSFER_BUFFER_SIZE:
assert nixl_config.buffer_size % NixlPipe.TRANSFER_BUFFER_SIZE == 0, (
f"Buffer size must be a multiple of {NixlPipe.TRANSFER_BUFFE... | or just don't pass in None here | """
self.nixl_config = nixl_config
self.side_channel = side_channel
if nixl_config.buffer_size > NixlPipe.TRANSFER_BUFFER_SIZE:
assert nixl_config.buffer_size % NixlPipe.TRANSFER_BUFFER_SIZE == 0, (
f"Buffer size must be a multiple of {NixlPipe.TRANSFER_BUFFE... | @@ -235,7 +235,12 @@ def __init__(
# allocator (should be initialized after self._buffer)
self._allocator = NixlBufferAllocator(self)
- self._agent = nixl_agent(str(nixl_config.role) + str(nixl_config.buffer_device))
+ self._agent = nixl_agent( | lmcache/v1/storage_backend/connector/nixl_connector_v2.py | 26 | Python | 0.071 | suggestion | 31 | 51 | 51 | false | [Misc] Allow passing nixl backends when creating nixl_agent in nixl_connector_v3 | 1,343 | LMCache/LMCache | 4,864 | Python | sammshen | jinqinn | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/nixl_connector_v2.py
PR Title: [Misc] Allow passing nixl backends when creating nixl_agent in nixl_connector_v3
BEFORE CODE: """
self.nixl_config = nixl_config
self.side_channel = side_channel
... | or just don't pass in None here | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
self.use_hot = config.local_cpu
self.lookup_server = lookup_server
self.memory_allocator = memory_allocator
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
self.cpu_lock = threading.Lock()
self.stream = torch.cuda.Stream()
... | this is to distinguish between workers right? | self.memory_allocator = memory_allocator
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
self.cpu_lock = threading.Lock()
self.stream = torch.cuda.Stream()
self.stats_monitor = LMCStatsMonitor.GetOrCreate()
self.layerwise = co... | @@ -67,6 +69,21 @@ def __init__(
# assumption: only one request is looked up at a time
# (only one worker per cache engine)
self.keys_in_request: List[CacheEngineKey] = []
+ self._setup_metrics()
+
+ def _setup_metrics(self):
+ Gauge(
+ "lmcache:local_cpu_hot_cache... | lmcache/v1/storage_backend/local_cpu_backend.py | 26 | Python | 0.071 | question | 46 | 51 | 51 | false | Add metrics to LocalCPUBackend by lambda | 1,344 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_cpu_backend.py
PR Title: Add metrics to LocalCPUBackend by lambda
BEFORE CODE: self.use_hot = config.local_cpu
self.lookup_server = lookup_server
self.memory_allocator = memory_allocator
self.lmc... | this is to distinguish between workers right? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
@staticmethod
def unregister_all_metrics():
collectors = list(REGISTRY._collector_to_names.keys())
for collector in collectors:
try:
REGISTRY.unregister(collector)
except KeyError:
pass
class PrometheusLogger:
_gauge_cls = prometheus_... | sorry where is this supposed to be called? | remote_cache_usage_bytes=self.remote_cache_usage_bytes,
local_storage_usage_bytes=self.local_storage_usage_bytes,
active_memory_objs_count=self.active_memory_objs_count,
pinned_memory_objs_count=self.pinned_memory_objs_count,
time_to_retrieve=time_to_retrieve,... | @@ -396,6 +397,15 @@ def GetOrCreate() -> "LMCStatsMonitor":
def DestroyInstance():
LMCStatsMonitor._instance = None
+ @staticmethod
+ def unregister_all_metrics(): | lmcache/observability.py | 26 | Python | 0.071 | question | 43 | 51 | 51 | false | Add metrics to LocalCPUBackend by lambda | 1,344 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/observability.py
PR Title: Add metrics to LocalCPUBackend by lambda
BEFORE CODE: @staticmethod
def unregister_all_metrics():
collectors = list(REGISTRY._collector_to_names.keys())
for collector in collectors:
try:
... | sorry where is this supposed to be called? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
self.cache_policy = get_cache_policy(config.cache_policy)
self.hot_cache = self.cache_policy.init_mutable_mapping()
self.use_hot = config.local_cpu
self.lookup_server = lookup_server
self.memory_allocator = memory_allocator
self.lmcache_worker = lmcache_worker
se... | Sorry, I'm not so familiar with the Prometheus python SDK. If we create gauges here, can they be automatically exposed to the prometheus metric endpoint? |
self.use_hot = config.local_cpu
self.lookup_server = lookup_server
self.memory_allocator = memory_allocator
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
self.cpu_lock = threading.Lock()
self.stream = torch.cuda.Stream()
... | @@ -67,6 +69,21 @@ def __init__(
# assumption: only one request is looked up at a time
# (only one worker per cache engine)
self.keys_in_request: List[CacheEngineKey] = []
+ self._setup_metrics()
+
+ def _setup_metrics(self):
+ Gauge( | lmcache/v1/storage_backend/local_cpu_backend.py | 26 | Python | 0.357 | question | 153 | 51 | 51 | false | Add metrics to LocalCPUBackend by lambda | 1,344 | LMCache/LMCache | 4,864 | Python | ApostaC | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_cpu_backend.py
PR Title: Add metrics to LocalCPUBackend by lambda
BEFORE CODE: self.cache_policy = get_cache_policy(config.cache_policy)
self.hot_cache = self.cache_policy.init_mutable_mapping()
self.us... | Sorry, I'm not so familiar with the Prometheus python SDK. If we create gauges here, can they be automatically exposed to the prometheus metric endpoint? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
self.cache_policy = get_cache_policy(config.cache_policy)
self.hot_cache = self.cache_policy.init_mutable_mapping()
self.use_hot = config.local_cpu
self.lookup_server = lookup_server
self.memory_allocator = memory_allocator
self.lmcache_worker = lmcache_worker
se... | Also, is it possible to add a helper function in `PrometheusLogger` for other modules to register metrics?
The goal of such a helper function is to make sure that the metric labels (right now include `model_id` and `worker_id`) can be correctly added. (See the function `PrometheusLogger._metadata_to_labels`) |
self.use_hot = config.local_cpu
self.lookup_server = lookup_server
self.memory_allocator = memory_allocator
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
self.cpu_lock = threading.Lock()
self.stream = torch.cuda.Stream()
... | @@ -67,6 +69,21 @@ def __init__(
# assumption: only one request is looked up at a time
# (only one worker per cache engine)
self.keys_in_request: List[CacheEngineKey] = []
+ self._setup_metrics()
+
+ def _setup_metrics(self):
+ Gauge( | lmcache/v1/storage_backend/local_cpu_backend.py | 26 | Python | 0.571 | refactor | 312 | 51 | 51 | false | Add metrics to LocalCPUBackend by lambda | 1,344 | LMCache/LMCache | 4,864 | Python | ApostaC | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_cpu_backend.py
PR Title: Add metrics to LocalCPUBackend by lambda
BEFORE CODE: self.cache_policy = get_cache_policy(config.cache_policy)
self.hot_cache = self.cache_policy.init_mutable_mapping()
self.us... | Also, is it possible to add a helper function in `PrometheusLogger` for other modules to register metrics?
The goal of such a helper function is to make sure that the metric labels (right now include `model_id` and `worker_id`) can be correctly added. (See the function `PrometheusLogger._metadata_to_labels`) | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
_instance = None
@staticmethod
def GetOrCreate(metadata: LMCacheEngineMetadata) -> "PrometheusLogger":
if PrometheusLogger._instance is None:
PrometheusLogger._instance = PrometheusLogger(metadata)
# assert PrometheusLogger._instance.metadata == metadata, \
# "Promet... | `GetInstanceIfCreate` sounds a little bit weird (I'm okay with it though)
Proposing some alternatives:
- `GetInstanceIfExists`
- `GetInstanceOrNone`
- `MaybeGetInstance`
|
_instance = None
@staticmethod
def GetOrCreate(metadata: LMCacheEngineMetadata) -> "PrometheusLogger":
if PrometheusLogger._instance is None:
PrometheusLogger._instance = PrometheusLogger(metadata)
# assert PrometheusLogger._instance.metadata == metadata, \
# "Promet... | @@ -910,6 +938,14 @@ def GetInstance() -> "PrometheusLogger":
)
return PrometheusLogger._instance
+ @staticmethod
+ def GetInstanceIfCreate() -> Optional["PrometheusLogger"]: | lmcache/observability.py | 26 | Python | 0.5 | suggestion | 177 | 51 | 51 | false | Add metrics to LocalCPUBackend by lambda | 1,344 | LMCache/LMCache | 4,864 | Python | ApostaC | maobaolong | Review this pull request change.
Language: Python
File: lmcache/observability.py
PR Title: Add metrics to LocalCPUBackend by lambda
BEFORE CODE:
_instance = None
@staticmethod
def GetOrCreate(metadata: LMCacheEngineMetadata) -> "PrometheusLogger":
if PrometheusLogger._instance is None:
... | `GetInstanceIfCreate` sounds a little bit weird (I'm okay with it though)
Proposing some alternatives:
- `GetInstanceIfExists`
- `GetInstanceOrNone`
- `MaybeGetInstance`
| You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
):
return 0
self._requests_priority[request.request_id] = request.priority
token_ids = request.prompt_token_ids
# If the request has multimodal hashes, apply them to the token ids
if request.mm_hashes:
# TODO(Jiayi): Optimize this
token_ids ... | QQ: is `request.priority` already stable enough in vLLM? | ####################
@_lmcache_nvtx_annotate
def get_num_new_matched_tokens(
self,
request: "Request",
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
request (Request): the request object.
... | @@ -1008,6 +1022,8 @@ def get_num_new_matched_tokens(
):
return 0
+ self._requests_priority[request.request_id] = request.priority | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.357 | question | 56 | 51 | 51 | false | Priority based storing -- only store kv cache for high priority requests | 1,368 | LMCache/LMCache | 4,864 | Python | ApostaC | FerdinandZhong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Priority based storing -- only store kv cache for high priority requests
BEFORE CODE: ):
return 0
self._requests_priority[request.request_id] = request.priority
token_i... | QQ: is `request.priority` already stable enough in vLLM? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Disagg spec for the request
disagg_spec: Optional[DisaggSpec] = None
# Multimodal hashes and positions
mm_hashes: Optional[list[str]] = None
mm_positions: Optional[list["PlaceholderRange"]] = None
# The configs of the request, includes tags and other configs
request_configs: Optional[di... | Why do we need to put the skip_save flag in the request tracker? This class is meant to save the request states, such as token ids/block ids/number of cached tokens.
An alternative way is to check the priority in `build_connector_meta` and directly calculate and pass the `skip_save` flag in the `build_connector_meta... | # FIXME: need to check whether the block ids will be changed after
# preemption
allocated_block_ids: list[int]
# The number of tokens that has been saved
num_saved_tokens: int = 0
# Disagg spec for the request
disagg_spec: Optional[DisaggSpec] = None
# Multimodal hashes and pos... | @@ -135,13 +135,18 @@ class RequestTracker:
# Whether the request is in decode phase
is_decode_phase = False
+ # Whether the request cache should be saved
+ skip_save: bool = False
+
@_lmcache_nvtx_annotate
@staticmethod
def from_new_request(
lmcache_config: LMCacheEngineConfig... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.571 | performance | 332 | 51 | 51 | false | Priority based storing -- only store kv cache for high priority requests | 1,368 | LMCache/LMCache | 4,864 | Python | ApostaC | FerdinandZhong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Priority based storing -- only store kv cache for high priority requests
BEFORE CODE:
# Disagg spec for the request
disagg_spec: Optional[DisaggSpec] = None
# Multimodal hashes and positions
... | Why do we need to put the skip_save flag in the request tracker? This class is meant to save the request states, such as token ids/block ids/number of cached tokens.
An alternative way is to check the priority in `build_connector_meta` and directly calculate and pass the `skip_save` flag in the `build_connector_meta... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
"""
Check for external KV cache hit.
Args:
request (Request): the request object.
num_computed_tokens (int): the number of locally
computed tokens for this request
Returns:
the number of tokens that can be loaded from the
... | The `request.priority` is not exists in vllm `0.8.5`, so this PR make LMCache break backward-compatible.
@ApostaC @FerdinandZhong @yoo-kumaneko | ####################
@_lmcache_nvtx_annotate
def get_num_new_matched_tokens(
self,
request: "Request",
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
request (Request): the request object.
... | @@ -1036,6 +1044,8 @@ def get_num_new_matched_tokens(
):
return 0
+ self._requests_priority[request.request_id] = request.priority | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.5 | performance | 147 | 51 | 51 | false | Priority based storing -- only store kv cache for high priority requests | 1,368 | LMCache/LMCache | 4,864 | Python | maobaolong | FerdinandZhong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Priority based storing -- only store kv cache for high priority requests
BEFORE CODE: """
Check for external KV cache hit.
Args:
request (Request): the request object.
... | The `request.priority` is not exists in vllm `0.8.5`, so this PR make LMCache break backward-compatible.
@ApostaC @FerdinandZhong @yoo-kumaneko | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Whether the scheduler allow us to load the tokens
can_load: bool
@dataclass
class SaveSpec:
# Skip already saved tokens
skip_leading_tokens: int
# Whether the scheduler allow us to save the tokens
can_save: bool
@dataclass
class DisaggSpec:
req_id: str
receiver_info: NixlReceiverIn... | This is better than set a `tag_keys` into confiig.👍 |
@dataclass
class SaveSpec:
# Skip already saved tokens
skip_leading_tokens: int
# Whether the scheduler allow us to save the tokens
can_save: bool
@dataclass
class DisaggSpec:
req_id: str
receiver_info: NixlReceiverInfo
is_last_prefill: bool = False
num_transferred_tokens: int = 0
t... | @@ -91,21 +91,16 @@ class DisaggSpec:
tmp_disagg_tracker: dict[str, DisaggSpec] = {}
-def extract_tags(
- config: LMCacheEngineConfig, sampling_params: SamplingParams
-) -> OrderedDict:
- tags = None
- tag_keys = None
- if config.extra_config is not None and isinstance(config.extra_config, dict):
- ... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.429 | suggestion | 52 | 51 | 51 | false | [refact] use request_configs replace tags | 1,377 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [refact] use request_configs replace tags
BEFORE CODE: # Whether the scheduler allow us to load the tokens
can_load: bool
@dataclass
class SaveSpec:
# Skip already saved tokens
skip_leadin... | This is better than set a `tag_keys` into confiig.👍 | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Whether to discard partial chunks
self._discard_partial_chunks = (
vllm_config.kv_transfer_config.get_from_extra_config(
"discard_partial_chunks", False
)
or not config.save_unfull_chunk
)
self._lmcache_chunk_size = config.chunk_siz... | nice, now we can also debug the scheduler independently |
# Whether to discard partial chunks
self._discard_partial_chunks = (
vllm_config.kv_transfer_config.get_from_extra_config(
"discard_partial_chunks", False
)
or not config.save_unfull_chunk
)
self._lmcache_chunk_size = config.chunk_siz... | @@ -610,6 +612,9 @@ def __init__(
self.force_skip_save = bool(os.environ.get("LMCACHE_FORCE_SKIP_SAVE", False))
+ self.api_server = InternalAPIServer(self) | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.357 | bug | 55 | 51 | 51 | false | Improve internal api server to support multiple lmcache server | 1,387 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Improve internal api server to support multiple lmcache server
BEFORE CODE:
# Whether to discard partial chunks
self._discard_partial_chunks = (
vllm_config.kv_transfer_config.g... | nice, now we can also debug the scheduler independently | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# request_id -> full_token_ids
self._request_trackers: dict[str, RequestTracker] = {}
# Whether to discard partial chunks
self._discard_partial_chunks = (
vllm_config.kv_transfer_config.get_from_extra_config(
"discard_partial_chunks", False
)
... | nit: can you write a small comment here that activation of API server for this worker / scheduler connector component is decided inside the constructor or else it looks like we always create the `api_server` |
# request_id -> full_token_ids
self._request_trackers: dict[str, RequestTracker] = {}
# Whether to discard partial chunks
self._discard_partial_chunks = (
vllm_config.kv_transfer_config.get_from_extra_config(
"discard_partial_chunks", False
)
... | @@ -610,6 +612,9 @@ def __init__(
self.force_skip_save = bool(os.environ.get("LMCACHE_FORCE_SKIP_SAVE", False))
| lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.571 | nitpick | 207 | 51 | 51 | false | Improve internal api server to support multiple lmcache server | 1,387 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Improve internal api server to support multiple lmcache server
BEFORE CODE:
# request_id -> full_token_ids
self._request_trackers: dict[str, RequestTracker] = {}
# Whether to disca... | nit: can you write a small comment here that activation of API server for this worker / scheduler connector component is decided inside the constructor or else it looks like we always create the `api_server` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
import struct
MAX_KEY_LENGTH = 150
class Constants(IntEnum):
CLIENT_PUT = auto()
CLIENT_GET = auto()
CLIENT_EXIST = auto()
CLIENT_LIST = auto()
SERVER_SUCCESS = 200
SERVER_FAIL =... | I know this isn't part of this PR but do you mind breaking this out into 2 enums:
- ClientCommands
- ServerReturnCodes
As it currently stands, the enum is just a catchall for constants which is not good practice. | # SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
import struct
MAX_KEY_LENGTH = 150
class ClientCommand(IntEnum):
PUT = auto()
GET = auto()
EXIST = auto()
LIST = auto()
class ServerReturnCode(IntEnum):
# keep the same as HTTP stat... | @@ -1,16 +1,17 @@
# SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
+from enum import IntEnum, auto
import struct
MAX_KEY_LENGTH = 150
-class Constants:
- CLIENT_PUT = 1
- CLIENT_GET = 2
- CLIENT_EXIST = 3
- CLIENT_LIST = 4
+class Constants(IntEnum):
+ CLIENT_P... | lmcache/protocol.py | 17 | Python | 0.429 | suggestion | 219 | 42 | 42 | false | use enum auto instead of number literal | 1,390 | LMCache/LMCache | 4,864 | Python | hickeyma | mengbingrock | Review this pull request change.
Language: Python
File: lmcache/protocol.py
PR Title: use enum auto instead of number literal
BEFORE CODE: # SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
import struct
MAX_KEY_LENGTH = 150
class Constants(IntEnum):... | I know this isn't part of this PR but do you mind breaking this out into 2 enums:
- ClientCommands
- ServerReturnCodes
As it currently stands, the enum is just a catchall for constants which is not good practice. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Standard
from collections import deque
from contextlib import nullcontext
from dataclasses import dataclass
from enum import Enum, auto
from typing import List, Optional, Tuple, Union
import abc
import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmcache.... | Its probably better to start with `0` like current value. | # Standard
from collections import deque
from contextlib import nullcontext
from dataclasses import dataclass
from enum import Enum, auto
from typing import Any, List, Optional, Tuple, Union
import abc
import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmc... | @@ -24,26 +24,26 @@
class MemoryFormat(Enum):
- UNDEFINED = 0
+ UNDEFINED = auto() | lmcache/v1/memory_management.py | 26 | Python | 0.429 | suggestion | 57 | 51 | 51 | false | use enum auto instead of number literal | 1,390 | LMCache/LMCache | 4,864 | Python | hickeyma | mengbingrock | Review this pull request change.
Language: Python
File: lmcache/v1/memory_management.py
PR Title: use enum auto instead of number literal
BEFORE CODE: # Standard
from collections import deque
from contextlib import nullcontext
from dataclasses import dataclass
from enum import Enum, auto
from typing import List, Op... | Its probably better to start with `0` like current value. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
from dataclasses import dataclass
from enum import Enum, auto
from typing import List, Optional, Tuple, Union
import abc
import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmcache.logging import init_logger
from lmcache.observability import LMCStatsMonitor... | Comments should probably retain the specific value and not `auto()` | from dataclasses import dataclass
from enum import Enum, auto
from typing import Any, List, Optional, Tuple, Union
import abc
import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmcache.logging import init_logger
from lmcache.observability import LMCStatsMo... | @@ -24,26 +24,26 @@
class MemoryFormat(Enum):
- UNDEFINED = 0
+ UNDEFINED = auto()
"""[2, num_layers, num_tokens, hidden_dim]
"""
- # KV_BLOB = 1
- KV_2LTD = 1
+ # KV_BLOB = auto() | lmcache/v1/memory_management.py | 26 | Python | 0.429 | suggestion | 67 | 51 | 51 | false | use enum auto instead of number literal | 1,390 | LMCache/LMCache | 4,864 | Python | hickeyma | mengbingrock | Review this pull request change.
Language: Python
File: lmcache/v1/memory_management.py
PR Title: use enum auto instead of number literal
BEFORE CODE: from dataclasses import dataclass
from enum import Enum, auto
from typing import List, Optional, Tuple, Union
import abc
import ctypes
import math
import threading
... | Comments should probably retain the specific value and not `auto()` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmcache.logging import init_logger
from lmcache.observability import LMCStatsMonitor
from lmcache.utils import _lmcache_nvtx_annotate
import lmcache.c_ops as lmc_ops
logger = init_logger(__name__)
class... | Leave comment with specific value and not auto() | import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmcache.logging import init_logger
from lmcache.observability import LMCStatsMonitor
from lmcache.utils import _lmcache_nvtx_annotate
import lmcache.c_ops as lmc_ops
logger = init_logger(__name__)
class... | @@ -28,22 +28,22 @@ class MemoryFormat(Enum):
"""[2, num_layers, num_tokens, hidden_dim]
"""
# KV_BLOB = 1
- KV_2LTD = 1
+ KV_2LTD = auto()
"""[num_tokens, 2, hidden_dim]
"""
- # LAYER_KV_BLOB = 2
- KV_T2D = 2
+ # LAYER_KV_BLOB = auto() | lmcache/v1/memory_management.py | 26 | Python | 0.071 | suggestion | 48 | 51 | 51 | false | use enum auto instead of number literal | 1,390 | LMCache/LMCache | 4,864 | Python | hickeyma | mengbingrock | Review this pull request change.
Language: Python
File: lmcache/v1/memory_management.py
PR Title: use enum auto instead of number literal
BEFORE CODE: import ctypes
import math
import threading
# Third Party
import sortedcontainers
import torch
# First Party
from lmcache.logging import init_logger
from lmcache.ob... | Leave comment with specific value and not auto() | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
from typing import Optional
import struct
# Third Party
import torch
# First Party
from lmcache.utils import CacheEngineKey
from lmcache.v1.memory_management import MemoryFormat
MAX_KEY_LENGTH = 150
cl... | Sorry but I didn't mean to add `s` at end; do you mind changing to `ClientCommand`? | # SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
from typing import Optional
import struct
# Third Party
import torch
# First Party
from lmcache.utils import CacheEngineKey
from lmcache.v1.memory_management import MemoryFormat
MAX_KEY_LENGTH = 150
cl... | @@ -14,15 +15,17 @@
MAX_KEY_LENGTH = 150
-class Constants:
- CLIENT_PUT = 1
- CLIENT_GET = 2
- CLIENT_EXIST = 3
- CLIENT_LIST = 4
- CLIENT_HEALTH = 5
+class ClientCommands(IntEnum): | lmcache/v1/protocol.py | 18 | Python | 0.357 | question | 83 | 43 | 43 | false | use enum auto instead of number literal | 1,390 | LMCache/LMCache | 4,864 | Python | hickeyma | mengbingrock | Review this pull request change.
Language: Python
File: lmcache/v1/protocol.py
PR Title: use enum auto instead of number literal
BEFORE CODE: # SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
from typing import Optional
import struct
# Third Party
imp... | Sorry but I didn't mean to add `s` at end; do you mind changing to `ClientCommand`? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
from typing import Optional
import struct
# Third Party
import torch
# First Party
from lmcache.utils import CacheEngineKey
from lmcache.v1.memory_management import MemoryFormat
MAX_KEY_LENGTH = 150
cl... | Sorry but I didn't mean to add `s` at end; do you mind changing to `ServerReturnCode`? | # SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
from typing import Optional
import struct
# Third Party
import torch
# First Party
from lmcache.utils import CacheEngineKey
from lmcache.v1.memory_management import MemoryFormat
MAX_KEY_LENGTH = 150
cl... | @@ -14,15 +15,17 @@
MAX_KEY_LENGTH = 150
-class Constants:
- CLIENT_PUT = 1
- CLIENT_GET = 2
- CLIENT_EXIST = 3
- CLIENT_LIST = 4
- CLIENT_HEALTH = 5
+class ClientCommands(IntEnum):
+ PUT = auto()
+ GET = auto()
+ EXIST = auto()
+ LIST = auto()
+ HEALTH = auto()
- SERVER_SUCCESS ... | lmcache/v1/protocol.py | 26 | Python | 0.357 | question | 86 | 51 | 51 | false | use enum auto instead of number literal | 1,390 | LMCache/LMCache | 4,864 | Python | hickeyma | mengbingrock | Review this pull request change.
Language: Python
File: lmcache/v1/protocol.py
PR Title: use enum auto instead of number literal
BEFORE CODE: # SPDX-License-Identifier: Apache-2.0
# Standard
from dataclasses import dataclass
from enum import IntEnum, auto
from typing import Optional
import struct
# Third Party
imp... | Sorry but I didn't mean to add `s` at end; do you mind changing to `ServerReturnCode`? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
_CONFIG_DEFINITIONS: dict[str, dict[str, Any]] = {
# Basic configurations
"chunk_size": {"type": int, "default": 256, "env_converter": int},
"local_cpu": {
"type": bool,
"default": True,
"env_converter": lambda x: x.lower() in ["true", "1"],
},
"max_local_cpu_size": {"type": ... | @Shaoting-Feng As this code duplicated so much time, would you like to abstract this into a `bool_converter` ?
You can referent to `_to_int_list` | "chunk_size": {"type": int, "default": 256, "env_converter": int},
"local_cpu": {
"type": bool,
"default": True,
"env_converter": lambda x: x
if isinstance(x, bool)
else str(x).lower() in ["true", "1"],
},
"max_local_cpu_size": {"type": float, "default": 5.0, "env... | @@ -78,12 +78,16 @@ def _to_int_list(
"use_layerwise": {
"type": bool,
"default": False,
- "env_converter": lambda x: x.lower() in ["true", "1"],
+ "env_converter": lambda x: x | lmcache/v1/config.py | 26 | Python | 0.5 | refactor | 148 | 51 | 51 | false | [Bugfix][PD] Fix config setting | 1,391 | LMCache/LMCache | 4,864 | Python | maobaolong | Shaoting-Feng | Review this pull request change.
Language: Python
File: lmcache/v1/config.py
PR Title: [Bugfix][PD] Fix config setting
BEFORE CODE: _CONFIG_DEFINITIONS: dict[str, dict[str, Any]] = {
# Basic configurations
"chunk_size": {"type": int, "default": 256, "env_converter": int},
"local_cpu": {
"type": ... | @Shaoting-Feng As this code duplicated so much time, would you like to abstract this into a `bool_converter` ?
You can referent to `_to_int_list` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
from lmcache.v1.rpc_utils import get_zmq_rpc_path_lmcache
if TYPE_CHECKING:
# Third Party
from vllm.config import VllmConfig
logger = init_logger(__name__)
class LMCacheLookupClient(LookupClientInterface):
"""
ZMQ-based lookup client that communicates with a lookup server.
Related extra_config:... | Suggest to abstract this `use_mla` to `utils.py` and modified the duplicated code also | from lmcache.v1.lookup_client.abstract_client import LookupClientInterface
from lmcache.v1.rpc_utils import get_zmq_rpc_path_lmcache
if TYPE_CHECKING:
# Third Party
from vllm.config import VllmConfig
logger = init_logger(__name__)
class LMCacheLookupClient(LookupClientInterface):
"""
ZMQ-based looku... | @@ -39,11 +39,17 @@ def __init__(self, vllm_config: "VllmConfig", config: LMCacheEngineConfig):
"lmcache_rpc_port", 0
)
self.tensor_parallel_size = vllm_config.parallel_config.tensor_parallel_size
+ use_mla = ( | lmcache/v1/lookup_client/lmcache_lookup_client.py | 26 | Python | 0.429 | refactor | 86 | 51 | 51 | false | [bugfix][#1357] Unified the default value if use mla | 1,394 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/lookup_client/lmcache_lookup_client.py
PR Title: [bugfix][#1357] Unified the default value if use mla
BEFORE CODE: from lmcache.v1.rpc_utils import get_zmq_rpc_path_lmcache
if TYPE_CHECKING:
# Third Party
from vllm.config import VllmConfig
... | Suggest to abstract this `use_mla` to `utils.py` and modified the duplicated code also | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
from lmcache.v1.lookup_client.lmcache_lookup_client import (
LMCacheLookupClient,
)
return LMCacheLookupClient(vllm_config, config)
@staticmethod
def create_lookup_server(
lmcache_engine: LMCacheEngine,
vllm_config: "VllmConfig",
) -> Opt... | Suggest to abstract a util method `get_extra_config_value(config, default_value)`, this can helps to reduce changes line a lot. | from lmcache.v1.lookup_client.lmcache_lookup_client import (
LMCacheLookupClient,
)
return LMCacheLookupClient(vllm_config, config)
@staticmethod
def create_lookup_server(
lmcache_engine: LMCacheEngine,
vllm_config: "VllmConfig",
) -> Opt... | @@ -72,11 +72,14 @@ def create_lookup_server(
# Only create the KV lookup API server on worker rank 0
# when there are multiple workers and when not using external lookup client
create_lookup_server_only_on_worker_0_for_mla = ( | lmcache/v1/lookup_client/factory.py | 26 | Python | 0.571 | refactor | 127 | 51 | 51 | false | [bugfix][#1357] Unified the default value if use mla | 1,394 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/lookup_client/factory.py
PR Title: [bugfix][#1357] Unified the default value if use mla
BEFORE CODE: from lmcache.v1.lookup_client.lmcache_lookup_client import (
LMCacheLookupClient,
)
return LMCa... | Suggest to abstract a util method `get_extra_config_value(config, default_value)`, this can helps to reduce changes line a lot. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
from lmcache.v1.storage_backend.connector.nixl_utils import (
get_correct_nixl_device,
)
corrected_device = get_correct_nixl_device(
config.nixl_buffer_device,
metadata.worker_id,
)
... | nit:
```python
config.get_extra_config_value("save_only_first_rank") and metadata.use_mla
```
works here since None is falsy
metadata.use_mla and metadata.use_mla just looks weird | dtype=torch.uint8,
device=corrected_device,
)
nixl_cpu_mem_allocator = NixlCPUMemoryAllocator()
nixl_cpu_mem_allocator.init_nixl_memory_allocator(
buffer,
torch.Size(metadata.kv_shape),
... | @@ -1201,13 +1197,10 @@ def _Create_memory_allocator(
return CuFileMemoryAllocator(config.cufile_buffer_size * 1024**2)
max_local_cpu_size = config.max_local_cpu_size
- save_only_first_rank_default = True if metadata.use_mla else False
+ # save_only_first_rank only works when use m... | lmcache/v1/cache_engine.py | 26 | Python | 0.786 | nitpick | 188 | 51 | 51 | false | [bugfix][#1357] Unified the default value if use mla | 1,394 | LMCache/LMCache | 4,864 | Python | sammshen | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [bugfix][#1357] Unified the default value if use mla
BEFORE CODE: from lmcache.v1.storage_backend.connector.nixl_utils import (
get_correct_nixl_device,
)
... | nit:
```python
config.get_extra_config_value("save_only_first_rank") and metadata.use_mla
```
works here since None is falsy
metadata.use_mla and metadata.use_mla just looks weird | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
if path.is_file():
files = [path]
elif path.is_dir():
# Recursively find all .py and .sh files
for ext in ["*.py", "*.sh"]:
files.extend(path.rglob(ext))
for file in files:
self._launch_plugin(file)
def _should_skip_plugin(sel... | nit:
```python
logger.info(f"worker {self.worker_id} is skipping plugin {file}, which is only for worker ID {plugin_worker_id}")
``` | if path.is_file():
files = [path]
elif path.is_dir():
# Recursively find all .py and .sh files
for ext in ["*.py", "*.sh"]:
files.extend(path.rglob(ext))
for file in files:
self._launch_plugin(file)
def _should_skip_plugin(sel... | @@ -0,0 +1,148 @@
+# SPDX-License-Identifier: Apache-2.0
+# Standard
+from pathlib import Path
+import atexit
+import os
+import subprocess
+import threading
+
+# First Party
+from lmcache.logging import init_logger
+
+logger = init_logger(__name__)
+
+
+class PluginLauncher:
+ def __init__(self, config, role, worke... | lmcache/v1/plugin/plugin_launcher.py | 26 | Python | 0.714 | nitpick | 136 | 51 | 51 | false | [feat] Introduce a plugin framework | 1,401 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/plugin/plugin_launcher.py
PR Title: [feat] Introduce a plugin framework
BEFORE CODE: if path.is_file():
files = [path]
elif path.is_dir():
# Recursively find all .py and .sh files
for ext in ["*.py... | nit:
```python
logger.info(f"worker {self.worker_id} is skipping plugin {file}, which is only for worker ID {plugin_worker_id}")
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# SPDX-License-Identifier: Apache-2.0
# Standard
from typing import List, Optional, Tuple, no_type_check
import asyncio
import os
import struct
# Third Party
from redis.asyncio import from_url as redis_from_url
from redis.asyncio.sentinel import Sentinel
import redis
# First Party
from lmcache.logging import init_log... | Main Change #2: I had to change redis to batched and async in order to encounter cpu memory pressure. this is not the main purpose of the PR | # SPDX-License-Identifier: Apache-2.0
# Standard
from typing import List, Optional, Tuple, no_type_check
import asyncio
import inspect
import os
# Third Party
import redis
# First Party
from lmcache.logging import init_logger
from lmcache.utils import CacheEngineKey
from lmcache.v1.memory_management import MemoryObj
... | lmcache/v1/storage_backend/connector/redis_connector.py | 1 | Python | 0.571 | performance | 140 | 26 | 26 | false | [patch]: busy loop on local cpu full | 1,429 | LMCache/LMCache | 4,864 | Python | sammshen | sammshen | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/redis_connector.py
PR Title: [patch]: busy loop on local cpu full
BEFORE CODE: # SPDX-License-Identifier: Apache-2.0
# Standard
from typing import List, Optional, Tuple, no_type_check
import asyncio
import os
import struct
... | Main Change #2: I had to change redis to batched and async in order to encounter cpu memory pressure. this is not the main purpose of the PR | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. | |
self.lookup_server = lookup_server
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
self.worker_id = metadata.worker_id
self.nixl_offload_stream = torch.cuda.Stream()
@_lmcache_nvtx_annotate
def allocate(
self,
shape: t... | This line causes problems. Some allocator backends don't have `busy_loop` and will cause runtime error | self.lookup_server = lookup_server
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
self.worker_id = metadata.worker_id
self.nixl_offload_stream = torch.cuda.Stream()
@_lmcache_nvtx_annotate
def allocate(
self,
shape: t... | @@ -103,14 +103,17 @@ def allocate(
dtype: torch.dtype,
fmt: MemoryFormat = MemoryFormat.KV_2LTD,
eviction=True,
+ busy_loop=True,
) -> Optional[MemoryObj]:
"""
Allocate memory object with memory allocator.
Use LRU evictor if eviction is enabled.
... | lmcache/v1/storage_backend/storage_manager.py | 26 | Python | 0.714 | suggestion | 103 | 51 | 51 | false | [patch]: busy loop on local cpu full | 1,429 | LMCache/LMCache | 4,864 | Python | ApostaC | sammshen | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/storage_manager.py
PR Title: [patch]: busy loop on local cpu full
BEFORE CODE: self.lookup_server = lookup_server
self.lmcache_worker = lmcache_worker
self.instance_id = config.lmcache_instance_id
sel... | This line causes problems. Some allocator backends don't have `busy_loop` and will cause runtime error | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
if mla is enabled, default is 0;
- if mla_lookup_server_worker_id < 0, start lookup server on all workers
- if mla_lookup server_worker_id >= 0, start lookup server on the given worker
"""
def __init__(
self,
vllm_config: "VllmConfig",
):
metadata, config = c... | The loop for range is not suitable for `self.mla_lookup_server_worker_id >= 0` i guess,
Would you like to generate a list and add all rank into it? You can loop by use that list instead of the `range(rank)`, maybe this can make code clearer than before, WDYT? | - mla_lookup_server_worker_id:
is a flag to control whether to create lookup server only on one worker.
if mla is not enabled, default is -1;
if mla is enabled, default is 0;
- if mla_lookup_server_worker_id < 0, start lookup server on all workers
- if mla_lookup server_worke... | @@ -47,17 +51,17 @@ def __init__(
"lmcache_rpc_port", 0
)
self.tensor_parallel_size = vllm_config.parallel_config.tensor_parallel_size
- use_mla = mla_enabled(vllm_config.model_config)
- self.create_lookup_server_only_on_worker_0_for_mla = (
- config.get_extra_con... | lmcache/v1/lookup_client/lmcache_async_lookup_client.py | 26 | Python | 0.643 | question | 260 | 51 | 51 | false | support start lookup server on other rank | 1,466 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/lookup_client/lmcache_async_lookup_client.py
PR Title: support start lookup server on other rank
BEFORE CODE: if mla is enabled, default is 0;
- if mla_lookup_server_worker_id < 0, start lookup server on all workers
- if mla_... | The loop for range is not suitable for `self.mla_lookup_server_worker_id >= 0` i guess,
Would you like to generate a list and add all rank into it? You can loop by use that list instead of the `range(rank)`, maybe this can make code clearer than before, WDYT? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
local_cpu_backend,
dst_device,
lmcache_worker,
lookup_server,
)
backend_name = str(local_disk_backend)
storage_backends[backend_name] = local_disk_backend
if config.weka_path is not None:
weka_backend = WekaGdsBackend(config, loop, me... | Can we make this a single if instead of if-else?
```
if xxx:
create_dynamix_backends(xxx)
``` | lmcache_worker,
lookup_server,
)
backend_name = str(local_disk_backend)
storage_backends[backend_name] = local_disk_backend
if config.weka_path is not None:
weka_backend = WekaGdsBackend(config, loop, memory_allocator, dst_device)
# TODO(Serapheim): ... | @@ -198,17 +198,20 @@ def CreateStorageBackends(
backend_name = str(remote_backend)
storage_backends[backend_name] = remote_backend
- # Create dynamic backends from configuration
- create_dynamic_backends(
- config,
- metadata,
- loop,
- memory_allocator,
- l... | lmcache/v1/storage_backend/__init__.py | 26 | Python | 0.643 | question | 99 | 51 | 51 | false | [Bugfix][PD] Fix nixl initialization problem | 1,494 | LMCache/LMCache | 4,864 | Python | YaoJiayi | Shaoting-Feng | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/__init__.py
PR Title: [Bugfix][PD] Fix nixl initialization problem
BEFORE CODE: local_cpu_backend,
dst_device,
lmcache_worker,
lookup_server,
)
backend_name = str(local... | Can we make this a single if instead of if-else?
```
if xxx:
create_dynamix_backends(xxx)
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# change that in the future.
storage_backends[str(weka_backend)] = weka_backend
if config.gds_path is not None:
gds_backend = GdsBackend(config, loop, memory_allocator, dst_device)
storage_backends[str(gds_backend)] = gds_backend
if config.remote_url is not None:
remote_b... | Using a placeholder like `pass` is ok in test code but not here. I think the code below is cleaner and is easier to understand the logic.
```suggestion
if not config.enable_nixl or config.local_cpu:
# Create dynamic backends from configuration
create_dynamic_backends(
config,
... | if config.gds_path is not None:
gds_backend = GdsBackend(config, loop, memory_allocator, dst_device)
storage_backends[str(gds_backend)] = gds_backend
if config.remote_url is not None:
remote_backend = RemoteBackend(
config, metadata, loop, local_cpu_backend, dst_device, looku... | @@ -198,17 +198,20 @@ def CreateStorageBackends(
backend_name = str(remote_backend)
storage_backends[backend_name] = remote_backend
- # Create dynamic backends from configuration
- create_dynamic_backends(
- config,
- metadata,
- loop,
- memory_allocator,
- l... | lmcache/v1/storage_backend/__init__.py | 26 | Python | 0.929 | suggestion | 520 | 50 | 46 | false | [Bugfix][PD] Fix nixl initialization problem | 1,494 | LMCache/LMCache | 4,864 | Python | hickeyma | Shaoting-Feng | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/__init__.py
PR Title: [Bugfix][PD] Fix nixl initialization problem
BEFORE CODE: # change that in the future.
storage_backends[str(weka_backend)] = weka_backend
if config.gds_path is not None:
gds_backend =... | Using a placeholder like `pass` is ok in test code but not here. I think the code below is cleaner and is easier to understand the logic.
```suggestion
if not config.enable_nixl or config.local_cpu:
# Create dynamic backends from configuration
create_dynamic_backends(
config,
... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
)
self.use_layerwise = config.use_layerwise
self.num_layers = metadata.kv_shape[0]
if self.use_layerwise:
if config.enable_blending:
self.fmt = MemoryFormat.KV_2TD
else:
self.fmt = MemoryFormat.KV_T2D
# NOTE(ApostaC): ... | caught by integration tests ! :)
```text
(EngineCore_0 pid=67) File "/opt/venv/lib/python3.12/site-packages/lmcache/v1/cache_engine.py", line 173, in __init__
--
| (EngineCore_0 pid=67) self.force_store_wait = config.extra_config.get("force_store_wait", False)
| (EngineCore_0 pid=67) ... | )
self.use_layerwise = config.use_layerwise
self.num_layers = metadata.kv_shape[0]
if self.use_layerwise:
if config.enable_blending:
self.fmt = MemoryFormat.KV_2TD
else:
self.fmt = MemoryFormat.KV_T2D
# NOTE(ApostaC): ... | @@ -169,6 +169,12 @@ def __init__(
self.post_inited = False
+ # Whether to force store to wait if no CPU buffer is available
+ self.force_store_wait = (
+ config.extra_config and \
+ config.extra_config.get("force_store_wait", False
+ )
+ | lmcache/v1/cache_engine.py | 26 | Python | 1 | performance | 580 | 51 | 51 | false | [Misc][Core] add extra config `force_store_wait` to make sure store does not skip any requests | 1,499 | LMCache/LMCache | 4,864 | Python | sammshen | ApostaC | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [Misc][Core] add extra config `force_store_wait` to make sure store does not skip any requests
BEFORE CODE: )
self.use_layerwise = config.use_layerwise
self.num_layers = metadata.kv_shape[0]
... | caught by integration tests ! :)
```text
(EngineCore_0 pid=67) File "/opt/venv/lib/python3.12/site-packages/lmcache/v1/cache_engine.py", line 173, in __init__
--
| (EngineCore_0 pid=67) self.force_store_wait = config.extra_config.get("force_store_wait", False)
| (EngineCore_0 pid=67) ... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
raise ValueError(f"No adapter found for URL: {self.context.url}")
def CreateConnector(
url: str,
loop: asyncio.AbstractEventLoop,
local_cpu_backend: LocalCPUBackend,
config: Optional[LMCacheEngineConfig] = None,
metadata: Optional[LMCacheEngineMetadata] = None,
) -> InstrumentedRemoteConn... | `benchmark` --> `mock`? Because its a mock connector that can be used for bench marking. |
raise ValueError(f"No adapter found for URL: {self.context.url}")
def CreateConnector(
url: str,
loop: asyncio.AbstractEventLoop,
local_cpu_backend: LocalCPUBackend,
config: Optional[LMCacheEngineConfig] = None,
metadata: Optional[LMCacheEngineMetadata] = None,
) -> InstrumentedRemoteConn... | @@ -219,6 +219,7 @@ def CreateConnector(
- audit://host:port[?verify=true|false]
- fs://[host:port]/path
- s3://[bucket].s3express-[az_id].[region].amazonaws.com"
+ - benchmark://[capacity]/?peeking_latency=[ms]&read_throughput=[GB/s]&write_throughput=[GB/s] | lmcache/v1/storage_backend/connector/__init__.py | 26 | Python | 0.429 | suggestion | 88 | 51 | 51 | false | [Feat]: Mock Connector | 1,500 | LMCache/LMCache | 4,864 | Python | hickeyma | sammshen | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/__init__.py
PR Title: [Feat]: Mock Connector
BEFORE CODE:
raise ValueError(f"No adapter found for URL: {self.context.url}")
def CreateConnector(
url: str,
loop: asyncio.AbstractEventLoop,
local_cpu_ba... | `benchmark` --> `mock`? Because its a mock connector that can be used for bench marking. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
)
self.plugin_processes.append(proc)
logger.info(f"Launched plugin: {file} with {interpreter}")
# Start thread to capture output continuo... | Error reading plugin file {file}: {e} --> Error reading interpreter from plugin file {file} - using default interpreters: {e} | env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
)
self.plugin_processes.append(proc)
logger.info(f"Launched plugin: {file} with {interpreter}")
# Start thread to capture output continuo... | @@ -109,22 +110,32 @@ def _launch_plugin(self, file: Path):
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
+ interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
first_line = f.readline().strip()
... | lmcache/v1/plugin/plugin_launcher.py | 26 | Python | 0.5 | suggestion | 125 | 51 | 51 | false | [Improvement][PluginFramework] Add default interpreter from plugin script | 1,504 | LMCache/LMCache | 4,864 | Python | hickeyma | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/plugin/plugin_launcher.py
PR Title: [Improvement][PluginFramework] Add default interpreter from plugin script
BEFORE CODE: env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
tex... | Error reading plugin file {file}: {e} --> Error reading interpreter from plugin file {file} - using default interpreters: {e} | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
).start()
except Exception as e:
logger.error(f"Failed to launch plugin {file}: {e}")
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
... | ```suggestion
else:
raise ValueError(f"Plugin type {file.suffix} not supported ")
``` | ).start()
except Exception as e:
logger.error(f"Failed to launch plugin {file}: {e}")
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
... | @@ -109,22 +110,32 @@ def _launch_plugin(self, file: Path):
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
+ interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
first_line = f.readline().strip()
... | lmcache/v1/plugin/plugin_launcher.py | 26 | Python | 0.857 | suggestion | 116 | 51 | 51 | false | [Improvement][PluginFramework] Add default interpreter from plugin script | 1,504 | LMCache/LMCache | 4,864 | Python | hickeyma | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/plugin/plugin_launcher.py
PR Title: [Improvement][PluginFramework] Add default interpreter from plugin script
BEFORE CODE: ).start()
except Exception as e:
logger.error(f"Failed to launch plugin {file}: {e}")
def... | ```suggestion
else:
raise ValueError(f"Plugin type {file.suffix} not supported ")
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
first_line = f.readline().strip()
if first_line.startswith("#!"):
... | resolved_interp --> resolved_interpreter |
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
first_line = f.readline().strip()
if first_line.startswith("#!"):
... | @@ -109,22 +110,32 @@ def _launch_plugin(self, file: Path):
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
+ interpreters = []
try:
with open(file, "r", encoding="utf-8") as f:
first_line = f.readline().strip()
... | lmcache/v1/plugin/plugin_launcher.py | 26 | Python | 0.071 | suggestion | 40 | 51 | 51 | false | [Improvement][PluginFramework] Add default interpreter from plugin script | 1,504 | LMCache/LMCache | 4,864 | Python | hickeyma | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/plugin/plugin_launcher.py
PR Title: [Improvement][PluginFramework] Add default interpreter from plugin script
BEFORE CODE:
def _get_interpreter(self, file: Path) -> str:
"""Get interpreter from first line comment"""
interpreters... | resolved_interp --> resolved_interpreter | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def relative_time(df, start_time):
"""
Relative time to the start of the benchmark.
"""
df["request_start"] = df["request_start"] - start_time
df["request_end"] = df["request_end"] - start_time
df["ttft_time"] = df["request_start"] + df["ttft"]
def visualize_results(warmup_df, benchmark_df):... | nit: can we change the "Request TTFT" to "Request first token"? |
def relative_time(df, start_time):
"""
Relative time to the start of the benchmark.
"""
df["request_start"] = df["request_start"] - start_time
df["request_end"] = df["request_end"] - start_time
df["ttft_time"] = df["request_start"] + df["ttft"]
def visualize_results(warmup_df, benchmark_df):... | @@ -242,6 +312,55 @@ def repeat_prompts(prompts, repeat_count, mode: str):
)
+def relative_time(df, start_time):
+ """
+ Relative time to the start of the benchmark.
+ """
+ df["request_start"] = df["request_start"] - start_time
+ df["request_end"] = df["request_end"] - start_time
+ df["t... | benchmarks/long_doc_qa/long_doc_qa.py | 26 | Python | 0.286 | question | 63 | 51 | 51 | false | [Benchmark] Add completions and visualization to long doc qa | 1,519 | LMCache/LMCache | 4,864 | Python | ApostaC | sammshen | Review this pull request change.
Language: Python
File: benchmarks/long_doc_qa/long_doc_qa.py
PR Title: [Benchmark] Add completions and visualization to long doc qa
BEFORE CODE:
def relative_time(df, start_time):
"""
Relative time to the start of the benchmark.
"""
df["request_start"] = df["reques... | nit: can we change the "Request TTFT" to "Request first token"? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Adapted from
# https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_long_document_qa_throughput.py
"""
Commandline arguments:
--num-documents: The number of documents to sample prompts from.
... | Is this still the default method? | # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Adapted from
# https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_long_document_qa_throughput.py
"""
Commandline arguments:
--num-documents: The number of documents to sample prompts from.
... | @@ -17,11 +17,10 @@
(Optional, default: 2)
--repeat-mode: The mode to repeat prompts. The supported modes are:
- - 'random': shuffle the prompts randomly. (Default)
- - 'tile': the entire prompt list is repeated in sequence. (Potentially
- lowest cache hit)
+ ... | benchmarks/long_doc_qa/long_doc_qa.py | 20 | Python | 0.071 | question | 33 | 45 | 45 | false | [Benchmark]: Benchmarking Recommendation + Documentation | 1,534 | LMCache/LMCache | 4,864 | Python | Shaoting-Feng | sammshen | Review this pull request change.
Language: Python
File: benchmarks/long_doc_qa/long_doc_qa.py
PR Title: [Benchmark]: Benchmarking Recommendation + Documentation
BEFORE CODE: # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Adapted from
# https://github.com... | Is this still the default method? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
if actual_latency_gain < args.expected_latency_gain:
sys.exit(
f"ERROR: latency gain {actual_latency_gain:.2f}× < expected "
f"{args.expected_latency_gain:.2f}×"
)
def create_argument_parser():
parser = argparse.ArgumentParser(
description="B... | Could you please also add this field to `LMCache/.buildkite/configs/xx.yaml`? I know this is not ideal but currently the comprehensive tests use a lot of default values in the `long_doc_qa.py`. | if actual_latency_gain < args.expected_latency_gain:
sys.exit(
f"ERROR: latency gain {actual_latency_gain:.2f}× < expected "
f"{args.expected_latency_gain:.2f}×"
)
def create_argument_parser():
parser = argparse.ArgumentParser(
description="B... | @@ -500,7 +518,7 @@ def create_argument_parser():
parser.add_argument(
"--num-documents",
type=int,
- default=8,
+ default=20, | benchmarks/long_doc_qa/long_doc_qa.py | 26 | Python | 0.571 | performance | 193 | 51 | 51 | false | [Benchmark]: Benchmarking Recommendation + Documentation | 1,534 | LMCache/LMCache | 4,864 | Python | Shaoting-Feng | sammshen | Review this pull request change.
Language: Python
File: benchmarks/long_doc_qa/long_doc_qa.py
PR Title: [Benchmark]: Benchmarking Recommendation + Documentation
BEFORE CODE: if actual_latency_gain < args.expected_latency_gain:
sys.exit(
f"ERROR: latency gain {actual_latency_gain:... | Could you please also add this field to `LMCache/.buildkite/configs/xx.yaml`? I know this is not ideal but currently the comprehensive tests use a lot of default values in the `long_doc_qa.py`. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
self.worker_id,
self.chunk_hash,
)
)
return hash(
(
self.fmt,
self.model_name,
self.world_size,
self.worker_id,
self.chunk_hash,
"%".joi... | Does this string concat can be avoid if we just use self.tags equals other.tags? | self.worker_id,
self.chunk_hash,
)
)
return hash(
(
self.fmt,
self.model_name,
self.world_size,
self.worker_id,
self.chunk_hash,
"%".joi... | @@ -126,6 +126,32 @@ def __hash__(self):
)
)
+ def __eq__(self, other):
+ if not isinstance(other, CacheEngineKey):
+ return False
+
+ if type(self) is CacheEngineKey and type(other) is not CacheEngineKey:
+ return False
+
+ self_tags_str = (
+ ... | lmcache/utils.py | 26 | Python | 0.286 | question | 80 | 51 | 51 | false | [bugfix] fix not hit in LocalCpuBackend | 1,545 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/utils.py
PR Title: [bugfix] fix not hit in LocalCpuBackend
BEFORE CODE: self.worker_id,
self.chunk_hash,
)
)
return hash(
(
self.fmt,
... | Does this string concat can be avoid if we just use self.tags equals other.tags? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
with self.key_lock:
for key in keys:
metadata = self.key_dict.get(key.chunk_hash)
if metadata is None:
obj_list.append(None)
continue
dtype = metadata.dtype
shape = metadata.shape
... | file_indices could also use a generic name | with self.key_lock:
for key in keys:
metadata = self.key_dict.get(key.chunk_hash)
if metadata is None:
obj_list.append(None)
continue
dtype = metadata.dtype
shape = metadata.shape
... | @@ -357,7 +437,7 @@ async def file_to_gpu(
if not mem_indices:
return obj_list
- handle = self.agent.get_file_to_gpu_handle(mem_indices, file_indices)
+ handle = self.agent.get_storage_to_mem_handle(mem_indices, file_indices) | lmcache/v1/storage_backend/nixl_storage_backend.py | 26 | Python | 0.143 | suggestion | 42 | 51 | 51 | false | [Core] Support NIXL storage obj backend | 1,557 | LMCache/LMCache | 4,864 | Python | jgoldsch12 | tshmilnvidia | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/nixl_storage_backend.py
PR Title: [Core] Support NIXL storage obj backend
BEFORE CODE: with self.key_lock:
for key in keys:
metadata = self.key_dict.get(key.chunk_hash)
if metadata ... | file_indices could also use a generic name | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
Args:
req_meta (ReqMeta): the request metadata.
"""
self.requests.append(req_meta)
class LMCacheConnectorV1Impl:
def __init__(
self,
vllm_config: "VllmConfig",
role: KVConnectorRole,
parent: KVConnectorBase_V1,
):
self._parent = paren... | the `LMCacheEngineConfig` is not a single instance, other `LMCacheEngineConfig` object does not have the configs from vllm when they are inited in other position, I am not sure if this will cause any inconsistency issues. | Args:
req_meta (ReqMeta): the request metadata.
"""
self.requests.append(req_meta)
class LMCacheConnectorV1Impl:
def __init__(
self,
vllm_config: "VllmConfig",
role: KVConnectorRole,
parent: KVConnectorBase_V1,
):
self._parent = paren... | @@ -547,7 +547,24 @@ def __init__(
assert isinstance(config, LMCacheEngineConfig), (
"LMCache v1 configuration is should be passed for vLLM v1."
)
+ # Put the leading with "lmcache." and matched configs from
+ # vllm extra_config to the config
+ kv_connector_extra_con... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.643 | performance | 221 | 51 | 51 | false | [Improve] Support pass through the lmcache config from vllm extra config | 1,568 | LMCache/LMCache | 4,864 | Python | chunxiaozheng | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [Improve] Support pass through the lmcache config from vllm extra config
BEFORE CODE: Args:
req_meta (ReqMeta): the request metadata.
"""
self.requests.append(req_meta)
... | the `LMCacheEngineConfig` is not a single instance, other `LMCacheEngineConfig` object does not have the configs from vllm when they are inited in other position, I am not sure if this will cause any inconsistency issues. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
class LMCacheConnectorV1Impl:
def __init__(
self,
vllm_config: "VllmConfig",
role: KVConnectorRole,
parent: KVConnectorBase_V1,
):
self._parent = parent
self.kv_role = vllm_config.kv_transfer_config.kv_role
self.worker_count = vllm_config.parallel_config... | nit: the validation happens with `hasattr()` and not with with the function `validate_and_set_config_value()` |
class LMCacheConnectorV1Impl:
def __init__(
self,
vllm_config: "VllmConfig",
role: KVConnectorRole,
parent: KVConnectorBase_V1,
):
self._parent = parent
self.kv_role = vllm_config.kv_transfer_config.kv_role
self.worker_count = vllm_config.parallel_config... | @@ -547,7 +547,24 @@ def __init__(
assert isinstance(config, LMCacheEngineConfig), (
"LMCache v1 configuration is should be passed for vLLM v1."
)
+ # Put the leading with "lmcache." and matched configs from
+ # vllm extra_config to the config
+ kv_connector_extra_con... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.5 | nitpick | 109 | 51 | 51 | false | [Improve] Support pass through the lmcache config from vllm extra config | 1,568 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [Improve] Support pass through the lmcache config from vllm extra config
BEFORE CODE:
class LMCacheConnectorV1Impl:
def __init__(
self,
vllm_config: "VllmConfig",
role: KVConne... | nit: the validation happens with `hasattr()` and not with with the function `validate_and_set_config_value()` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def mock_up_broadcast_object_fn(a: Any, i: int) -> None:
raise NotImplementedError("Calling invalid broadcast object function")
#### Python equivalents of CPP CUDA functions ####
def alloc_pinned_numa_ptr(size: int) -> torch.Tensor:
# Create a 1D uint8 CPU tensor in pinned memory, as uint8 == 1 byte
# R... | I like this idea of having the same set of C functions implemented in Python for non-CUDA platforms.
Can we separate those to a new source file so that we can import it as:
```
if torch.cuda.is_available():
import lmcache.c_ops as lmc_ops
else:
import lmcache.mockup_ops as lmc_ops # could have differe... | # SPDX-License-Identifier: Apache-2.0
# Future
from __future__ import annotations
# Standard
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, List, Optional, Tuple
import asyncio
import hashlib
import threading
import traceback
try:
# Third Party
from nvtx import annotate # type: igno... | @@ -384,3 +384,26 @@ def mock_up_broadcast_fn(t: torch.Tensor, i: int) -> None:
def mock_up_broadcast_object_fn(a: Any, i: int) -> None:
raise NotImplementedError("Calling invalid broadcast object function")
+
+
+#### Python equivalents of CPP CUDA functions ####
+def alloc_pinned_numa_ptr(size: int) -> torch.... | lmcache/utils.py | 26 | Python | 1 | performance | 438 | 27 | 387 | false | [Test]: Enable unit tests to run on non-CUDA environment | 1,575 | LMCache/LMCache | 4,864 | Python | ApostaC | hickeyma | Review this pull request change.
Language: Python
File: lmcache/utils.py
PR Title: [Test]: Enable unit tests to run on non-CUDA environment
BEFORE CODE:
def mock_up_broadcast_object_fn(a: Any, i: int) -> None:
raise NotImplementedError("Calling invalid broadcast object function")
#### Python equivalents of C... | I like this idea of having the same set of C functions implemented in Python for non-CUDA platforms.
Can we separate those to a new source file so that we can import it as:
```
if torch.cuda.is_available():
import lmcache.c_ops as lmc_ops
else:
import lmcache.mockup_ops as lmc_ops # could have differe... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
request: "Request",
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
request (Request): the request object.
num_computed_tokens (int): the number of locally
computed tokens for this request
... | Can we have a helper function called `extract_mm_features` that returns `Optional[Tuple[mm_hashes, mm_positions]]`?
So the code will be like:
```python
if mm_features := extract_mm_features(request):
token_ids = torch.tensor(request.prompt_token_ids)
mm_hashes, mm_positions = mm_features
apply_mm_ha... | def get_num_new_matched_tokens(
self,
request: "Request",
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
request (Request): the request object.
num_computed_tokens (int): the number of locally
... | @@ -1055,23 +1058,27 @@ def get_num_new_matched_tokens(
token_ids = request.prompt_token_ids
# If the request has multimodal hashes, apply them to the token ids
- if request.mm_features:
+ if request.mm_features or request.mm_hashes: | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.857 | refactor | 349 | 51 | 51 | false | [Core] Add backward compatibility with vllm | 1,583 | LMCache/LMCache | 4,864 | Python | ApostaC | Shaoting-Feng | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [Core] Add backward compatibility with vllm
BEFORE CODE: request: "Request",
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
A... | Can we have a helper function called `extract_mm_features` that returns `Optional[Tuple[mm_hashes, mm_positions]]`?
So the code will be like:
```python
if mm_features := extract_mm_features(request):
token_ids = torch.tensor(request.prompt_token_ids)
mm_hashes, mm_positions = mm_features
apply_mm_ha... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
request (Request): the request object.
num_computed_tokens (int): the number of locally
computed tokens for this request
Returns:
the... | ```suggestion
mm_hashes, mm_positions = extract_mm_features(request)
if mm_hashes and mm_positions:
``` | ###################
# Scheduler side APIs
####################
@_lmcache_nvtx_annotate
def get_num_new_matched_tokens(
self,
request: "Request",
num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
... | @@ -1055,23 +1052,17 @@ def get_num_new_matched_tokens(
token_ids = request.prompt_token_ids
# If the request has multimodal hashes, apply them to the token ids
- if request.mm_features:
+ mm_features = extract_mm_features(request)
+ if mm_features[0] != [] and mm_features[1] !=... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.786 | refactor | 122 | 51 | 51 | false | [Core] Add backward compatibility with vllm | 1,583 | LMCache/LMCache | 4,864 | Python | ApostaC | Shaoting-Feng | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: [Core] Add backward compatibility with vllm
BEFORE CODE: num_computed_tokens: int,
) -> Optional[int]:
"""
Check for external KV cache hit.
Args:
request (Re... | ```suggestion
mm_hashes, mm_positions = extract_mm_features(request)
if mm_hashes and mm_positions:
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
echo \"Model ${model} is available on ${port}\"
"; then
echo "OpenAI API server did not start"
docker logs $cid
return 1
fi
}
run_lmcache_vllmopenai_container() {
local docker="$1"
local vllm="$2"
local cfg_name="$3"
LOGFILE="/tmp/build_${BUILD_ID}_${cfg_name}.lo... | We can also mount the entire config directory so that it doesn’t have to be only for the MLA test. | echo \"Model ${model} is available on ${port}\"
"; then
echo "OpenAI API server did not start"
docker logs $cid
return 1
fi
}
run_lmcache_vllmopenai_container() {
local docker="$1"
local vllm="$2"
local cfg_name="$3"
LOGFILE="/tmp/build_${BUILD_ID}_${cfg_name}.lo... | @@ -111,16 +111,18 @@ run_lmcache_vllmopenai_container() {
local cfg_name="$3"
LOGFILE="/tmp/build_${BUILD_ID}_${cfg_name}.log"
- # Pick the GPU with the largest free memory
- source "$ORIG_DIR/.buildkite/scripts/pick-free-gpu.sh" ""
+ # Pick the GPUs based on config
+ gpu_count=$(yq -r '.docker... | .buildkite/scripts/vllm-integration-tests.sh | 26 | Shell | 0.214 | suggestion | 98 | 51 | 51 | false | Update vLLM options for MLA test | 1,588 | LMCache/LMCache | 4,864 | Python | Shaoting-Feng | yoo-kumaneko | Review this pull request change.
Language: Shell
File: .buildkite/scripts/vllm-integration-tests.sh
PR Title: Update vLLM options for MLA test
BEFORE CODE: echo \"Model ${model} is available on ${port}\"
"; then
echo "OpenAI API server did not start"
docker logs $cid
return 1
... | We can also mount the entire config directory so that it doesn’t have to be only for the MLA test. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
assert len(self.kv_caches) > 0
kvcaches = list(self.kv_caches.values())
attn_metadata = forward_context.attn_metadata
if attn_metadata is None:
logger.debug("In connector.start_load_kv, but the attn_metadata is None")
return
assert self.lmcache_engine is... | Using `kvcaches[0][0].device` looks messy. Can we have a `self.device` somewhere? | """Get vLLM version information.
Returns:
str: vLLM version string
"""
return VLLM_VERSION
@_lmcache_nvtx_annotate
def _init_kv_caches_from_forward_context(self, forward_context: "ForwardContext"):
for layer_name in forward_context.no_compile_layers:
... | @@ -822,7 +840,7 @@ def start_load_kv(self, forward_context: "ForwardContext", **kwargs) -> None:
tokens = request.token_ids
# TODO: have a pre-allocated buffer to hold the slot_mappings
- slot_mapping = request.slot_mapping.cuda()
+ slot_mapping = request.slot_mapping.... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.357 | performance | 81 | 51 | 51 | false | Add XPU support to LMCache for CPU/disk offloading | 1,593 | LMCache/LMCache | 4,864 | Python | YaoJiayi | zhenwei-intel | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Add XPU support to LMCache for CPU/disk offloading
BEFORE CODE: assert len(self.kv_caches) > 0
kvcaches = list(self.kv_caches.values())
attn_metadata = forward_context.attn_metadata... | Using `kvcaches[0][0].device` looks messy. Can we have a `self.device` somewhere? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
)
# Third Party
from vllm.attention.backends.flashmla import FlashMLAMetadata
from vllm.attention.backends.mla.common import MLACommonMetadata
from vllm.config import (
CacheConfig,
ModelConfig,
ParallelConfig,
SchedulerConfig,
)
from vllm.sequence import IntermediateTensors
from vllm.utils import ... | Would be great if you make this device check by vLLM way:
```
from vllm.platforms import current_platform
if current_platform.is_cuda_alike():
``` | )
# Third Party
from vllm.attention.backends.flashmla import FlashMLAMetadata
from vllm.attention.backends.mla.common import MLACommonMetadata
from vllm.config import (
CacheConfig,
ModelConfig,
ParallelConfig,
SchedulerConfig,
)
from vllm.platforms import current_platform
from vllm.sequence import... | @@ -49,7 +49,8 @@
logger = init_logger(__name__)
-LMCACHE_CUDA_STREAM = torch.cuda.Stream()
+if torch.cuda.is_available(): | lmcache/integration/vllm/vllm_adapter.py | 26 | Python | 0.714 | suggestion | 146 | 51 | 51 | false | Add XPU support to LMCache for CPU/disk offloading | 1,593 | LMCache/LMCache | 4,864 | Python | KuntaiDu | zhenwei-intel | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_adapter.py
PR Title: Add XPU support to LMCache for CPU/disk offloading
BEFORE CODE: )
# Third Party
from vllm.attention.backends.flashmla import FlashMLAMetadata
from vllm.attention.backends.mla.common import MLACommonMetada... | Would be great if you make this device check by vLLM way:
```
from vllm.platforms import current_platform
if current_platform.is_cuda_alike():
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
kvcaches = list(self.kv_caches.values())
attn_metadata = forward_context.attn_metadata
if attn_metadata is None:
logger.debug("In connector.start_load_kv, but the attn_metadata is None")
return
assert self.lmcache_engine is not None
self.lmcache_engine.... | this should not be the cause of the performance degradation. `.cuda()` should call `.to()` under the hood | assert len(self.kv_caches) > 0
kvcaches = list(self.kv_caches.values())
attn_metadata = forward_context.attn_metadata
if attn_metadata is None:
logger.debug("In connector.start_load_kv, but the attn_metadata is None")
return
assert self.lmcache_engine is... | @@ -863,7 +884,7 @@ def start_load_kv(self, forward_context: "ForwardContext", **kwargs) -> None:
tokens = request.token_ids
# TODO: have a pre-allocated buffer to hold the slot_mappings
- slot_mapping = request.slot_mapping.cuda()
+ slot_mapping = request.slot_mapping.... | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.714 | performance | 105 | 51 | 51 | false | Add XPU support to LMCache for CPU/disk offloading | 1,593 | LMCache/LMCache | 4,864 | Python | sammshen | zhenwei-intel | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Add XPU support to LMCache for CPU/disk offloading
BEFORE CODE: kvcaches = list(self.kv_caches.values())
attn_metadata = forward_context.attn_metadata
if attn_metadata is None:
... | this should not be the cause of the performance degradation. `.cuda()` should call `.to()` under the hood | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def _get_file_path(
self, key: CacheEngineKey, generate_tmp_path=False
) -> Tuple[Path, Optional[Path]]:
"""Get file path and tmp path for the given key"""
# If there's only one path, use it directly
if len(self.base_paths) == 1:
base_path = self.base_paths[0]
... | This method didn't need the second value of the tuple which returned by `_get_file_path ` | def _get_base_path(self, key: CacheEngineKey) -> Path:
"""Get file base path for the given key"""
if len(self.base_paths) == 1:
base_path = self.base_paths[0]
else:
# Calculate hash value and modulo to select path
hash_val = abs(key.chunk_hash)
... | @@ -69,21 +82,28 @@ def _get_file_path(self, key: CacheEngineKey) -> Path:
base_path = self.base_paths[idx]
key_path = key.to_string().replace("/", "-") + ".data"
- return base_path / key_path
+ data_path = base_path / key_path
+ tmp_path = None
+ if generate_tmp_path... | lmcache/v1/storage_backend/connector/fs_connector.py | 26 | Python | 0.357 | suggestion | 89 | 51 | 51 | false | support tmp path in FSConnector | 1,603 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/fs_connector.py
PR Title: support tmp path in FSConnector
BEFORE CODE: def _get_file_path(
self, key: CacheEngineKey, generate_tmp_path=False
) -> Tuple[Path, Optional[Path]]:
"""Get file path and tm... | This method didn't need the second value of the tuple which returned by `_get_file_path ` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
)
else:
user_session = UserSession(user_config, self.use_sharegpt)
self.sessions.append(user_session)
return user_session
def _remove_finished_sessions(self):
sessions_to_remove = [s for s in self.sessions if s.finished]
if len(sessions_to_remove) > 0... | In my opinion, we should not add this additional condition. We use this script to benchmark different engines. If an engine performs poorly and requests finish slowly, adding this condition will cause the QPS to decrease, making the workload lighter than for an engine that performs well. This would result in an unfair ... |
def _create_user_session(self):
self.user_id += 1
user_config = UserConfig.new_user_config(self.user_id, self.workload_config)
if self.use_sharegpt:
user_session = UserSession(
user_config, self.use_sharegpt, self.sharegpt_data[self.user_id]
)
... | @@ -426,7 +426,8 @@ def step(self, timestamp: float, executor: RequestExecutor):
if self.start_time is None:
self.start_time = timestamp
- if timestamp - self.last_user_join > self.gap_between_users:
+ # New user session only joins when meets time interval and active user count is ... | benchmarks/multi_round_qa/multi-round-qa.py | 26 | Python | 0.5 | performance | 331 | 51 | 51 | false | [Benchmark] configure user stacking at high qps in multi-round-qa | 1,609 | LMCache/LMCache | 4,864 | Python | Shaoting-Feng | yuezhu1 | Review this pull request change.
Language: Python
File: benchmarks/multi_round_qa/multi-round-qa.py
PR Title: [Benchmark] configure user stacking at high qps in multi-round-qa
BEFORE CODE: )
else:
user_session = UserSession(user_config, self.use_sharegpt)
self.sessions.append... | In my opinion, we should not add this additional condition. We use this script to benchmark different engines. If an engine performs poorly and requests finish slowly, adding this condition will cause the QPS to decrease, making the workload lighter than for an engine that performs well. This would result in an unfair ... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
send_path: str,
done_event: threading.Event,
):
"""
Upload a file to S3.
"""
headers = HttpHeaders()
headers.add("Host", self.s3_endpoint)
req = HttpRequest("PUT", self._format_safe_path(key_str), headers)
done = {"err": None, "status": None}... | Why do we have a `_put` and another `put`? | headers.add("Host", self.s3_endpoint)
req = HttpRequest("PUT", self._format_safe_path(key_str), headers)
done = {"err": None, "status": None}
def on_done(error=None, status_code=None, **kwargs):
done["err"] = error
done["status"] = status_code
if d... | @@ -464,7 +542,7 @@ def on_done(error=None, status_code=None, **kwargs):
on_done=on_done,
)
- async def put(self, key: CacheEngineKey, memory_obj: MemoryObj):
+ async def _put(self, key: CacheEngineKey, memory_obj: MemoryObj): | lmcache/v1/storage_backend/connector/s3_connector.py | 26 | Python | 0.214 | question | 42 | 51 | 51 | false | [feat]: working async s3 | 1,614 | LMCache/LMCache | 4,864 | Python | YaoJiayi | sammshen | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/s3_connector.py
PR Title: [feat]: working async s3
BEFORE CODE: send_path: str,
done_event: threading.Event,
):
"""
Upload a file to S3.
"""
headers = HttpHeaders()
... | Why do we have a `_put` and another `put`? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
priority=Priorities.PREFETCH,
)
async def list(self) -> List[str]:
raise NotImplementedError
def support_ping(self) -> bool:
return False
# TODO(Jiayi): This needs to be implemented.
async def ping(self) -> int:
raise NotImplementedError
def support_ba... | We can move the mhm-related cleanup code to the memory allocator | # SPDX-License-Identifier: Apache-2.0
# Standard
from enum import IntEnum, auto
from functools import partial
from typing import List, Optional
from urllib.parse import quote as url_quote
import asyncio
import ctypes
import mmap
import os
import tempfile
# Third Party
from awscrt import auth, io, s3
from awscrt.http i... | @@ -514,6 +659,12 @@ def support_batched_get(self) -> bool:
return True
async def close(self):
- for shm in self.shms:
- shm.close()
- shm.unlink()
+ await self.pq_executor.shutdown(wait=True)
+ # let python GC clean up mmap inodes
+ for mm in self.adhoc... | lmcache/v1/storage_backend/connector/s3_connector.py | 26 | Python | 0.429 | suggestion | 64 | 27 | 654 | false | [feat]: working async s3 | 1,614 | LMCache/LMCache | 4,864 | Python | YaoJiayi | sammshen | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/s3_connector.py
PR Title: [feat]: working async s3
BEFORE CODE: priority=Priorities.PREFETCH,
)
async def list(self) -> List[str]:
raise NotImplementedError
def support_ping(self) -> bo... | We can move the mhm-related cleanup code to the memory allocator | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
hash_buf,
offset_buf,
lookup_id_buf,
request_configs_buf,
]
else:
tokens_buf = self.encoder.encode(token_ids)
msg_buf = [
tokens_buf,
lookup_id_buf,
request_configs... | why don't we set socket timeouts once at initiatialization? | for start, end, key in self.token_database.process_tokens(
token_ids, make_key=False
):
hashes.append(key)
offsets.append(end - start)
hash_buf = self.encoder.encode(hashes)
offset_buf = self.encoder.encode(offsets)
... | @@ -128,16 +129,71 @@ def lookup(
request_configs_buf,
]
- for i in range(ranks):
- self.sockets[i].send_multipart(msg_buf, copy=False)
-
results = []
- # TODO(Jiayi): we can use zmq poll to optimize a bit
+ max_retries = self.config.lookup_max_re... | lmcache/v1/lookup_client/lmcache_lookup_client.py | 26 | Python | 0.214 | question | 60 | 51 | 51 | false | [Improve] Add timeout for lookup client | 1,615 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/lookup_client/lmcache_lookup_client.py
PR Title: [Improve] Add timeout for lookup client
BEFORE CODE: hash_buf,
offset_buf,
lookup_id_buf,
request_configs_buf,
]
el... | why don't we set socket timeouts once at initiatialization? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def preprocess_attention_qk(self, q: torch.Tensor, k: torch.Tensor, attn_layer):
"""
Qwen3-specific Q/K preprocessing - apply q_norm and k_norm
Args:
q: Query tensor
k: Key tensor
attn_layer: Attention layer
Returns:
tuple: (processe... | Can this function be fused with rotary embedding kernel? | # SPDX-License-Identifier: Apache-2.0
# First Party
from lmcache.v1.compute.models.base import LMCBaseModel
class LMCQwen3Model(LMCBaseModel):
def _process_qkv(self, q, k, v, layer):
"""Process QKV tensors for Qwen3 model with q_norm and k_norm layers."""
# Qwen3 has q_norm and k_norm layers
... | @@ -0,0 +1,75 @@
+# SPDX-License-Identifier: Apache-2.0
+# Third Party
+import torch
+
+# Local
+from .base import LMCBaseModel
+
+# TODO(Jiayi): A few things need to be tested/supported:
+# TP, PP, Multimodal
+
+
+class LMCQwen3Model(LMCBaseModel):
+ """LMC Qwen3 model implementation, inheriting from base model cla... | lmcache/v1/compute/models/qwen3.py | 26 | Python | 0.286 | question | 56 | 29 | 25 | false | [Model] Add Qwen3 model support for CacheBlend | 1,633 | LMCache/LMCache | 4,864 | Python | YaoJiayi | junl666 | Review this pull request change.
Language: Python
File: lmcache/v1/compute/models/qwen3.py
PR Title: [Model] Add Qwen3 model support for CacheBlend
BEFORE CODE:
def preprocess_attention_qk(self, q: torch.Tensor, k: torch.Tensor, attn_layer):
"""
Qwen3-specific Q/K preprocessing - apply q_norm a... | Can this function be fused with rotary embedding kernel? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
k: torch.Tensor,
v: torch.Tensor,
residual: torch.Tensor,
layer_id: int,
attn_output: Optional[torch.Tensor],
attn_metadata: LMCAttnMetadata,
):
logger.debug(f"Blender is processing KV for layer {layer_id}")
old_k, old_v = self.gpu_connector.get_kv(lay... | Is it a good idea to move the `rotary_emb` into `preprocess_attention_qk` | k: torch.Tensor,
v: torch.Tensor,
residual: torch.Tensor,
layer_id: int,
attn_output: Optional[torch.Tensor],
attn_metadata: LMCAttnMetadata,
):
logger.debug(f"Blender is processing KV for layer {layer_id}")
old_k, old_v = self.gpu_connector.get_kv(lay... | @@ -83,6 +83,10 @@ def process_qkv(
)
layer = self.layerwise_model.vllm_model.model.layers[layer_id]
attn_layer = layer.self_attn
+
+ # Use model-specific Q/K preprocessing method
+ q, k = self.layerwise_model.preprocess_attention_qk(q, k, attn_layer)
+
q, k = attn_... | lmcache/v1/compute/blend/blender.py | 26 | Python | 0.429 | suggestion | 73 | 51 | 51 | false | [Model] Add Qwen3 model support for CacheBlend | 1,633 | LMCache/LMCache | 4,864 | Python | YaoJiayi | junl666 | Review this pull request change.
Language: Python
File: lmcache/v1/compute/blend/blender.py
PR Title: [Model] Add Qwen3 model support for CacheBlend
BEFORE CODE: k: torch.Tensor,
v: torch.Tensor,
residual: torch.Tensor,
layer_id: int,
attn_output: Optional[torch.Tensor],
... | Is it a good idea to move the `rotary_emb` into `preprocess_attention_qk` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# First Party
from lmcache.config import LMCacheEngineMetadata
from lmcache.utils import CacheEngineKey
from lmcache.v1.config import LMCacheEngineConfig
from lmcache.v1.lookup_server import LookupServerInterface
from lmcache.v1.memory_management import (
MemoryAllocatorInterface,
MemoryFormat,
MemoryObj,
)... | Let's not include this for the default `StorageBackendInterface` since those parameters are pretty tightly coupled to each backend's implementation.
How about we create a base class for all the dynamic loading storage backend? Something like:
```python
class ExternalStorageBackend(StorageBackendInterface): # I'm ... | # First Party
from lmcache.config import LMCacheEngineMetadata
from lmcache.utils import CacheEngineKey
from lmcache.v1.config import LMCacheEngineConfig
from lmcache.v1.memory_management import (
MemoryAllocatorInterface,
MemoryFormat,
MemoryObj,
)
if TYPE_CHECKING:
# First Party
from lmcache.v1.s... | @@ -10,18 +11,28 @@
from lmcache.config import LMCacheEngineMetadata
from lmcache.utils import CacheEngineKey
from lmcache.v1.config import LMCacheEngineConfig
+from lmcache.v1.lookup_server import LookupServerInterface
from lmcache.v1.memory_management import (
MemoryAllocatorInterface,
MemoryFormat,
... | lmcache/v1/storage_backend/abstract_backend.py | 26 | Python | 0.857 | performance | 878 | 51 | 51 | false | [Backend][BugFix]: Refactor StorageBackendInterface constructor to enable dynamic loading | 1,636 | LMCache/LMCache | 4,864 | Python | ApostaC | hickeyma | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/abstract_backend.py
PR Title: [Backend][BugFix]: Refactor StorageBackendInterface constructor to enable dynamic loading
BEFORE CODE: # First Party
from lmcache.config import LMCacheEngineMetadata
from lmcache.utils import CacheEngine... | Let's not include this for the default `StorageBackendInterface` since those parameters are pretty tightly coupled to each backend's implementation.
How about we create a base class for all the dynamic loading storage backend? Something like:
```python
class ExternalStorageBackend(StorageBackendInterface): # I'm ... | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
module_path = config.extra_config.get(
f"external_backend.{backend_name}.module_path"
)
class_name = config.extra_config.get(
f"external_backend.{backend_name}.class_name"
)
if not module_path or not class_name:
... | @hickeyma As this PR modified this, and as the lmc_external_log_backend is a demonstrate project for `external backend framework`, would you like to modify the https://github.com/opendataio/lmc_external_log_backend/blob/main/lmc_external_log_backend/lmc_external_log_backend.py also? | f"external_backend.{backend_name}.module_path"
)
class_name = config.extra_config.get(
f"external_backend.{backend_name}.class_name"
)
if not module_path or not class_name:
logger.warning(
f"Backend {bac... | @@ -79,12 +79,12 @@ def create_dynamic_backends(
# Create the backend instance
backend_instance = backend_class( | lmcache/v1/storage_backend/__init__.py | 26 | Python | 0.571 | question | 283 | 51 | 51 | false | [Backend][BugFix]: Refactor StorageBackendInterface constructor to enable dynamic loading | 1,636 | LMCache/LMCache | 4,864 | Python | maobaolong | hickeyma | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/__init__.py
PR Title: [Backend][BugFix]: Refactor StorageBackendInterface constructor to enable dynamic loading
BEFORE CODE: module_path = config.extra_config.get(
f"external_backend.{backend_name}.module_... | @hickeyma As this PR modified this, and as the lmc_external_log_backend is a demonstrate project for `external backend framework`, would you like to modify the https://github.com/opendataio/lmc_external_log_backend/blob/main/lmc_external_log_backend/lmc_external_log_backend.py also? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def insert_put_task(self, key: CacheEngineKey):
with self.put_lock:
self.put_tasks.append(key)
def exists_in_put_tasks(self, key: CacheEngineKey) -> bool:
with self.put_lock:
return key in self.put_tasks
def close(self):
# Gracefully shut down the executor
... | Why these changes are necessary? @hickeyma | def insert_put_task(self, key: CacheEngineKey):
with self.put_lock:
self.put_tasks.append(key)
def exists_in_put_tasks(self, key: CacheEngineKey) -> bool:
with self.put_lock:
return key in self.put_tasks
def close(self):
# Gracefully shut down the executor
... | @@ -101,7 +101,7 @@ def __init__(
dst_device: str = "cuda",
lmcache_worker: Optional["LMCacheWorker"] = None,
):
- super().__init__(dst_device) | lmcache/v1/storage_backend/local_disk_backend.py | 26 | Python | 0.143 | question | 44 | 51 | 51 | false | [Backend][BugFix]: Refactor StorageBackendInterface constructor to enable dynamic loading | 1,636 | LMCache/LMCache | 4,864 | Python | maobaolong | hickeyma | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_disk_backend.py
PR Title: [Backend][BugFix]: Refactor StorageBackendInterface constructor to enable dynamic loading
BEFORE CODE: def insert_put_task(self, key: CacheEngineKey):
with self.put_lock:
self.p... | Why these changes are necessary? @hickeyma | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
metadata: Optional[LMCacheEngineMetadata] = None,
) -> MemoryAllocatorInterface:
cpu_size = config.max_local_cpu_size
if metadata is not None:
# save_only_first_rank only works when use mla
save_only_first_rank = (
config.get_extra_config_value("save_... | we should not always override the cpu size right, only in the save_only_first_rank case maybe | def initialize_allocator(
self,
config: LMCacheEngineConfig,
metadata: Optional[LMCacheEngineMetadata] = None,
) -> MemoryAllocatorInterface:
cpu_size = config.max_local_cpu_size
if metadata is not None:
# save_only_first_rank only works when use mla
... | @@ -266,17 +319,18 @@ def initialize_allocator(
if save_only_first_rank and metadata.is_first_rank():
# Only the first rank will save the cache,
- # so we need to set it lager than other ranks
- cpu_size = (
- config.extra_config.get("firs... | lmcache/v1/storage_backend/local_cpu_backend.py | 26 | Python | 0.286 | suggestion | 93 | 51 | 51 | false | Support adjust the effective memory by system available memory and reserve size | 1,708 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_cpu_backend.py
PR Title: Support adjust the effective memory by system available memory and reserve size
BEFORE CODE: metadata: Optional[LMCacheEngineMetadata] = None,
) -> MemoryAllocatorInterface:
cpu_size... | we should not always override the cpu size right, only in the save_only_first_rank case maybe | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
The effective CPU memory size in GB
"""
save_only_first_rank = (
metadata is not None
and config.get_extra_config_value("save_only_first_rank", metadata.use_mla)
and metadata.use_mla
)
if not save_only_first_rank:
# Do not adju... | `effective_cpu_size` may be 0, should an exception or assert be thrown in this situation? | The effective CPU memory size in GB
"""
save_only_first_rank = (
metadata is not None
and config.get_extra_config_value("save_only_first_rank", metadata.use_mla)
and metadata.use_mla
)
if not save_only_first_rank:
# Do not adju... | @@ -250,6 +250,61 @@ def remove(self, key: CacheEngineKey, force: bool = True) -> bool:
# other backends might still (temporarily) hold the memory object.
return True
+ def _calculate_effective_cpu_size(
+ self,
+ configured_cpu_size: float,
+ config: LMCacheEngineConfig,
+ ... | lmcache/v1/storage_backend/local_cpu_backend.py | 26 | Python | 0.571 | question | 89 | 51 | 51 | false | Support adjust the effective memory by system available memory and reserve size | 1,708 | LMCache/LMCache | 4,864 | Python | chunxiaozheng | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/local_cpu_backend.py
PR Title: Support adjust the effective memory by system available memory and reserve size
BEFORE CODE: The effective CPU memory size in GB
"""
save_only_first_rank = (
met... | `effective_cpu_size` may be 0, should an exception or assert be thrown in this situation? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
keys=keys,
memory_objs=memory_objs,
location=location,
)
return num_tokens
@_lmcache_nvtx_annotate
def lookup_unpin(self, lookup_ids: list[str]) -> None:
for lookup_id in lookup_ids:
if lookup_id in self.lookup_pins:
self.... | hi @maobaolong, why delete the broadcast here? | keys=keys,
memory_objs=memory_objs,
location=location,
)
return num_tokens
@_lmcache_nvtx_annotate
def lookup_unpin(self, lookup_ids: list[str]) -> None:
for lookup_id in lookup_ids:
if lookup_id in self.lookup_pins:
self.... | @@ -980,11 +980,9 @@ def clear(
if self.save_only_first_rank:
if self.metadata.is_first_rank():
num_removed = self._clear(tokens, locations, request_configs)
- self.broadcast_object_fn(num_removed, self.metadata.first_rank) | lmcache/v1/cache_engine.py | 26 | Python | 0.214 | question | 46 | 51 | 51 | false | [core]Add cache/clear api to internal_api_server | 1,711 | LMCache/LMCache | 4,864 | Python | chunxiaozheng | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [core]Add cache/clear api to internal_api_server
BEFORE CODE: keys=keys,
memory_objs=memory_objs,
location=location,
)
return num_tokens
@_lmcache_nvtx_annotate
... | hi @maobaolong, why delete the broadcast here? | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# TODO(Jiayi): we can use zmq poll to optimize a bit
for i in range(ranks):
resp = self.sockets[i].recv()
result = int.from_bytes(resp, "big")
results.append(result)
except zmq.Again:
logger.error(f"Timeout occurred for rank {i... | need to consult req_status before | results = []
try:
for i in range(ranks):
self.sockets[i].send_multipart(msg_buf, copy=False)
# TODO(Jiayi): we can use zmq poll to optimize a bit
for i in range(ranks):
resp = self.sockets[i].recv()
result = int.from_by... | @@ -163,7 +173,13 @@ def lookup(
# NOTE: it is possible that the number of hit tokens is different
# across TP ranks, so we can use the minimum value as the
# number of hit tokens. | lmcache/v1/lookup_client/lmcache_lookup_client.py | 26 | Python | 0.071 | suggestion | 33 | 51 | 51 | false | [Core] Remove `lookup_id` and use `req_id` instead | 1,729 | LMCache/LMCache | 4,864 | Python | sammshen | YaoJiayi | Review this pull request change.
Language: Python
File: lmcache/v1/lookup_client/lmcache_lookup_client.py
PR Title: [Core] Remove `lookup_id` and use `req_id` instead
BEFORE CODE:
# TODO(Jiayi): we can use zmq poll to optimize a bit
for i in range(ranks):
resp = self.sockets... | need to consult req_status before | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# TODO: extend this class to implement caching policies and eviction policies
class StorageManager:
"""
The StorageManager is responsible for managing the storage backends.
"""
def __init__(
self,
config: LMCacheEngineConfig,
metadata: LMCacheEngineMetadata,
event_manage... | can we create a helper
```python
is_cuda_worker(metadata):
return metadata.role != "scheduler" and torch.cuda.is_available()
``` | # TODO: extend this class to implement caching policies and eviction policies
class StorageManager:
"""
The StorageManager is responsible for managing the storage backends.
"""
def __init__(
self,
config: LMCacheEngineConfig,
metadata: LMCacheEngineMetadata,
event_manage... | @@ -195,7 +195,10 @@ def __init__(
)
self.thread.start()
- if torch.cuda.is_available():
+ # For scheduler role, always use CPU device
+ if metadata.role == "scheduler": | lmcache/v1/storage_backend/storage_manager.py | 26 | Python | 0.714 | refactor | 139 | 51 | 51 | false | Add a LMCacheBypassLookupClient to support lookup without communicate to any worker | 1,733 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/storage_manager.py
PR Title: Add a LMCacheBypassLookupClient to support lookup without communicate to any worker
BEFORE CODE: # TODO: extend this class to implement caching policies and eviction policies
class StorageManager:
"""... | can we create a helper
```python
is_cuda_worker(metadata):
return metadata.role != "scheduler" and torch.cuda.is_available()
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
args=(self.loop,),
name="storage-manger-event-loop",
)
self.thread.start()
# For scheduler role, always use CPU device
if is_cuda_worker(metadata):
dst_device = "cuda"
else:
dst_device = "cpu"
self.storage_backends: Ordered... | super nit (sorry):
```python
self.allocator_backend = None
if metadata.role != "scheduler":
self.allocator_backend = self._get_allocator_backend(config)
``` | args=(self.loop,),
name="storage-manger-event-loop",
)
self.thread.start()
# For scheduler role, always use CPU device
if is_cuda_worker(metadata):
dst_device = "cuda"
else:
dst_device = "cpu"
self.storage_backends: Ordered... | @@ -214,7 +215,11 @@ def __init__(
self.enable_pd = config.enable_pd
- self.allocator_backend = self._get_allocator_backend(config)
+ self.allocator_backend = ( | lmcache/v1/storage_backend/storage_manager.py | 26 | Python | 0.714 | nitpick | 171 | 51 | 51 | false | Add a LMCacheBypassLookupClient to support lookup without communicate to any worker | 1,733 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/storage_manager.py
PR Title: Add a LMCacheBypassLookupClient to support lookup without communicate to any worker
BEFORE CODE: args=(self.loop,),
name="storage-manger-event-loop",
)
self.thread.... | super nit (sorry):
```python
self.allocator_backend = None
if metadata.role != "scheduler":
self.allocator_backend = self._get_allocator_backend(config)
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
use_gpu=use_gpu,
chunk_size=chunk_size,
dtype=kv_dtype,
device=device,
)
tpg = get_tp_group()
else:
vllm_gpu_connector = VLLMPagedMemGPUConnectorV2(
hidden_dim_size,
num_layer,
use_gpu=use... | super nit (sorry):
```python
if role == "scheduler" and lmcache_config.enable_scheduler_bypass_lookup:
assert (
engine.save_only_first_rank
or lmcache_config.get_extra_config_value(
"remote_enable_mla_worker_id_as0", metadata.use_mla
)
)
``` | use_gpu=use_gpu,
chunk_size=chunk_size,
dtype=kv_dtype,
device=device,
)
tpg = get_tp_group()
else:
vllm_gpu_connector = VLLMPagedMemGPUConnectorV2(
hidden_dim_size,
num_layer,
use_gpu=use... | @@ -553,7 +560,19 @@ def _init_lmcache_engine(
tpg.broadcast,
tpg.broadcast_object,
)
-
+ assert ( | lmcache/integration/vllm/vllm_v1_adapter.py | 26 | Python | 0.857 | performance | 317 | 51 | 51 | false | Add a LMCacheBypassLookupClient to support lookup without communicate to any worker | 1,733 | LMCache/LMCache | 4,864 | Python | sammshen | maobaolong | Review this pull request change.
Language: Python
File: lmcache/integration/vllm/vllm_v1_adapter.py
PR Title: Add a LMCacheBypassLookupClient to support lookup without communicate to any worker
BEFORE CODE: use_gpu=use_gpu,
chunk_size=chunk_size,
dtype=kv_dtype,
... | super nit (sorry):
```python
if role == "scheduler" and lmcache_config.enable_scheduler_bypass_lookup:
assert (
engine.save_only_first_rank
or lmcache_config.get_extra_config_value(
"remote_enable_mla_worker_id_as0", metadata.use_mla
)
)
``` | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
# Deserialize metadata and allocate memory
metadata = RemoteMetadata.deserialize(md_buffer)
memory_obj = self.local_cpu_backend.allocate(
metadata.shape, metadata.dtype, metadata.fmt
)
else:
... | It will be graceful if you could care about the tail chunk case. |
# Deserialize metadata and allocate memory
metadata = RemoteMetadata.deserialize(md_buffer)
memory_obj = self.local_cpu_backend.allocate(
metadata.shape, metadata.dtype, metadata.fmt
)
else:
... | @@ -142,13 +156,25 @@ async def get(self, key: CacheEngineKey) -> Optional[MemoryObj]:
# Read the actual data into allocated memory
buffer = memory_obj.byte_array
- num_read = await f.readinto(buffer)
if self.save_chunk_meta:
+ num_r... | lmcache/v1/storage_backend/connector/fs_connector.py | 26 | Python | 0.286 | suggestion | 64 | 51 | 51 | false | [FSConnector] support read ahead in FSConnector | 1,771 | LMCache/LMCache | 4,864 | Python | maobaolong | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/connector/fs_connector.py
PR Title: [FSConnector] support read ahead in FSConnector
BEFORE CODE:
# Deserialize metadata and allocate memory
metadata = RemoteMetadata.deserialize(md_buffer)
... | It will be graceful if you could care about the tail chunk case. | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
keys = []
for chunk_info in chunk_info_iterator:
assert isinstance(chunk_info[2], CacheEngineKey)
chunk_info_list.append(chunk_info)
keys.append(chunk_info[2])
batched_contains_res = self.storage_manager.batched... | thanks for cleaning up the prev_end logic | continue
return res
else:
chunk_info_list = []
keys = []
for chunk_info in chunk_info_iterator:
assert isinstance(chunk_info[2], CacheEngineKey)
chunk_info_list.append(chunk_in... | @@ -713,29 +711,44 @@ def lookup(
found = True
if found:
if pin:
+ assert lookup_id is not None, (
+ "lookup_id is required when pin is True"
+ )
... | lmcache/v1/cache_engine.py | 26 | Python | 0.071 | suggestion | 41 | 51 | 51 | false | [core] add batched_contains interface | 1,778 | LMCache/LMCache | 4,864 | Python | sammshen | chunxiaozheng | Review this pull request change.
Language: Python
File: lmcache/v1/cache_engine.py
PR Title: [core] add batched_contains interface
BEFORE CODE: keys = []
for chunk_info in chunk_info_iterator:
assert isinstance(chunk_info[2], CacheEngineKey)
ch... | thanks for cleaning up the prev_end logic | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
def layerwise_batched_get(
self,
keys: List[List[CacheEngineKey]],
location: Optional[str] = None,
) -> Generator[Future, None, None]:
"""
Non-blocking function to get the memory objects into the storages
in a layerwise manner.
Do not store if the same obj... | Thanks for the fix @DongDongJu . instead of checking for if the serializer is None here, can we always intialize it (not only in the async codepath). Layerwise codepath also needs async serializer because multiple layers loading at once shoudl ensure that we don't deadlock the cpu as well | def layerwise_batched_get(
self,
keys: List[List[CacheEngineKey]],
location: Optional[str] = None,
) -> Generator[Future, None, None]:
"""
Non-blocking function to get the memory objects into the storages
in a layerwise manner.
Do not store if the same obj... | @@ -436,20 +437,15 @@ def layerwise_batched_get(
"""
if location is None:
location = "LocalCPUBackend"
-
for keys_multi_chunk in keys:
# Retrieve all chunks for one layer
backend = self.storage_backends[location]
# TODO(Jiayi): need to make ... | lmcache/v1/storage_backend/storage_manager.py | 26 | Python | 0.643 | bug | 289 | 51 | 51 | false | Guarded the async serializer usage so we only wrap the backend load c… | 1,795 | LMCache/LMCache | 4,864 | Python | sammshen | DongDongJu | Review this pull request change.
Language: Python
File: lmcache/v1/storage_backend/storage_manager.py
PR Title: Guarded the async serializer usage so we only wrap the backend load c…
BEFORE CODE: def layerwise_batched_get(
self,
keys: List[List[CacheEngineKey]],
location: Optional[str] =... | Thanks for the fix @DongDongJu . instead of checking for if the serializer is None here, can we always intialize it (not only in the async codepath). Layerwise codepath also needs async serializer because multiple layers loading at once shoudl ensure that we don't deadlock the cpu as well | You are an expert software engineer and code reviewer. Review pull requests carefully for correctness, bugs, security issues, performance problems, maintainability, and unnecessary complexity. Only raise actionable issues supported by the code. Explain why an issue matters and suggest a concrete fix when appropriate. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.