[Misc] Remove old now unsupported max_num_partial_prefills and max_long_partial_prefills (#49244)
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
This commit is contained in:
@@ -66,7 +66,6 @@ def test_worker_apply_lora(qwen3_lora_files):
|
|||||||
runner_type="generate",
|
runner_type="generate",
|
||||||
max_num_batched_tokens=32,
|
max_num_batched_tokens=32,
|
||||||
max_num_seqs=32,
|
max_num_seqs=32,
|
||||||
max_num_partial_prefills=32,
|
|
||||||
),
|
),
|
||||||
device_config=DeviceConfig(DEVICE_TYPE),
|
device_config=DeviceConfig(DEVICE_TYPE),
|
||||||
cache_config=CacheConfig(
|
cache_config=CacheConfig(
|
||||||
|
|||||||
@@ -67,16 +67,6 @@ class SchedulerConfig:
|
|||||||
In real usage, this should be set in `EngineArgs.create_engine_config`.
|
In real usage, this should be set in `EngineArgs.create_engine_config`.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
max_num_partial_prefills: int = Field(default=1, ge=1)
|
|
||||||
"""For chunked prefill, the maximum number of sequences that can be
|
|
||||||
partially prefilled concurrently."""
|
|
||||||
|
|
||||||
max_long_partial_prefills: int = Field(default=1, ge=1)
|
|
||||||
"""For chunked prefill, the maximum number of prompts longer than
|
|
||||||
long_prefill_token_threshold that will be prefilled concurrently. Setting
|
|
||||||
this less than max_num_partial_prefills will allow shorter prompts to jump
|
|
||||||
the queue in front of longer prompts in some cases, improving latency."""
|
|
||||||
|
|
||||||
long_prefill_token_threshold: int = Field(default=0, ge=0)
|
long_prefill_token_threshold: int = Field(default=0, ge=0)
|
||||||
"""For chunked prefill, a request is considered long if the prompt is
|
"""For chunked prefill, a request is considered long if the prompt is
|
||||||
longer than this number of tokens. 0 disables the cap (default)."""
|
longer than this number of tokens. 0 disables the cap (default)."""
|
||||||
@@ -254,19 +244,6 @@ class SchedulerConfig:
|
|||||||
self.max_num_batched_tokens,
|
self.max_num_batched_tokens,
|
||||||
)
|
)
|
||||||
|
|
||||||
if self.max_num_partial_prefills > 1:
|
|
||||||
if self.long_prefill_token_threshold == 0:
|
|
||||||
self.long_prefill_token_threshold = int(max_model_len * 0.04)
|
|
||||||
|
|
||||||
logger.info(
|
|
||||||
"Concurrent partial prefills enabled with "
|
|
||||||
"max_num_partial_prefills=%d, max_long_partial_prefills=%d, "
|
|
||||||
"long_prefill_token_threshold=%d",
|
|
||||||
self.max_num_partial_prefills,
|
|
||||||
self.max_long_partial_prefills,
|
|
||||||
self.long_prefill_token_threshold,
|
|
||||||
)
|
|
||||||
|
|
||||||
self.verify_max_model_len(max_model_len)
|
self.verify_max_model_len(max_model_len)
|
||||||
|
|
||||||
def verify_max_model_len(self, max_model_len: int) -> Self:
|
def verify_max_model_len(self, max_model_len: int) -> Self:
|
||||||
@@ -298,24 +275,11 @@ class SchedulerConfig:
|
|||||||
self.max_num_seqs * max_model_len,
|
self.max_num_seqs * max_model_len,
|
||||||
)
|
)
|
||||||
|
|
||||||
if self.max_num_partial_prefills > 1:
|
if self.long_prefill_token_threshold > max_model_len:
|
||||||
if not self.enable_chunked_prefill:
|
|
||||||
raise ValueError(
|
|
||||||
"Chunked prefill must be enabled to set "
|
|
||||||
"max_num_partial_prefills > 1."
|
|
||||||
)
|
|
||||||
|
|
||||||
if self.long_prefill_token_threshold > max_model_len:
|
|
||||||
raise ValueError(
|
|
||||||
"long_prefill_token_threshold "
|
|
||||||
f"({self.long_prefill_token_threshold}) cannot be greater "
|
|
||||||
f"than the max_model_len ({max_model_len})."
|
|
||||||
)
|
|
||||||
|
|
||||||
if self.max_long_partial_prefills > self.max_num_partial_prefills:
|
|
||||||
raise ValueError(
|
raise ValueError(
|
||||||
f"{self.max_long_partial_prefills=} must be less than or equal to "
|
"long_prefill_token_threshold "
|
||||||
f"{self.max_num_partial_prefills=}."
|
f"({self.long_prefill_token_threshold}) cannot be greater "
|
||||||
|
f"than the max_model_len ({max_model_len})."
|
||||||
)
|
)
|
||||||
|
|
||||||
return self
|
return self
|
||||||
|
|||||||
@@ -527,8 +527,6 @@ class EngineArgs:
|
|||||||
kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
|
kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
|
||||||
max_num_batched_tokens: int | None = None
|
max_num_batched_tokens: int | None = None
|
||||||
max_num_scheduled_tokens: int | None = None
|
max_num_scheduled_tokens: int | None = None
|
||||||
max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
|
|
||||||
max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
|
|
||||||
long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
|
long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
|
||||||
max_num_seqs: int | None = None
|
max_num_seqs: int | None = None
|
||||||
max_logprobs: int = ModelConfig.max_logprobs
|
max_logprobs: int = ModelConfig.max_logprobs
|
||||||
@@ -1440,13 +1438,6 @@ class EngineArgs:
|
|||||||
"default": None,
|
"default": None,
|
||||||
},
|
},
|
||||||
)
|
)
|
||||||
scheduler_group.add_argument(
|
|
||||||
"--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
|
|
||||||
)
|
|
||||||
scheduler_group.add_argument(
|
|
||||||
"--max-long-partial-prefills",
|
|
||||||
**scheduler_kwargs["max_long_partial_prefills"],
|
|
||||||
)
|
|
||||||
scheduler_group.add_argument(
|
scheduler_group.add_argument(
|
||||||
"--long-prefill-token-threshold",
|
"--long-prefill-token-threshold",
|
||||||
**scheduler_kwargs["long_prefill_token_threshold"],
|
**scheduler_kwargs["long_prefill_token_threshold"],
|
||||||
@@ -2190,8 +2181,6 @@ class EngineArgs:
|
|||||||
is_encoder_decoder=model_config.is_encoder_decoder,
|
is_encoder_decoder=model_config.is_encoder_decoder,
|
||||||
policy=self.scheduling_policy,
|
policy=self.scheduling_policy,
|
||||||
scheduler_cls=self.scheduler_cls,
|
scheduler_cls=self.scheduler_cls,
|
||||||
max_num_partial_prefills=self.max_num_partial_prefills,
|
|
||||||
max_long_partial_prefills=self.max_long_partial_prefills,
|
|
||||||
long_prefill_token_threshold=self.long_prefill_token_threshold,
|
long_prefill_token_threshold=self.long_prefill_token_threshold,
|
||||||
scheduler_reserve_full_isl=self.scheduler_reserve_full_isl,
|
scheduler_reserve_full_isl=self.scheduler_reserve_full_isl,
|
||||||
watermark=self.watermark,
|
watermark=self.watermark,
|
||||||
@@ -2402,14 +2391,6 @@ class EngineArgs:
|
|||||||
|
|
||||||
def _check_feature_supported(self):
|
def _check_feature_supported(self):
|
||||||
"""Raise an error if the feature is not supported."""
|
"""Raise an error if the feature is not supported."""
|
||||||
# No Concurrent Partial Prefills so far.
|
|
||||||
if (
|
|
||||||
self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills
|
|
||||||
or self.max_long_partial_prefills
|
|
||||||
!= SchedulerConfig.max_long_partial_prefills
|
|
||||||
):
|
|
||||||
_raise_unsupported_error(feature_name="Concurrent Partial Prefill")
|
|
||||||
|
|
||||||
if self.pipeline_parallel_size > 1:
|
if self.pipeline_parallel_size > 1:
|
||||||
supports_pp = getattr(
|
supports_pp = getattr(
|
||||||
self.distributed_executor_backend, "supports_pp", False
|
self.distributed_executor_backend, "supports_pp", False
|
||||||
|
|||||||
Reference in New Issue
Block a user