[2/N] Chunked prefill data update (#3538)

This commit is contained in:
SangBin Cho
2024-03-28 10:06:01 -07:00
committed by GitHub
parent ce567a2926
commit b51c1cc9d2
11 changed files with 272 additions and 76 deletions
+15 -5
View File
@@ -62,6 +62,7 @@ class EngineArgs:
image_input_shape: Optional[str] = None
image_feature_size: Optional[int] = None
scheduler_delay_factor: float = 0.0
enable_chunked_prefill: bool = False
def __post_init__(self):
if self.tokenizer is None:
@@ -356,6 +357,12 @@ class EngineArgs:
default=EngineArgs.scheduler_delay_factor,
help='Apply a delay (of delay factor multiplied by previous'
'prompt latency) before scheduling next prompt.')
parser.add_argument(
'--enable-chunked-prefill',
type=bool,
default=False,
help='If True, the prefill requests can be chunked based on the '
'max_num_batched_tokens')
return parser
@classmethod
@@ -394,11 +401,14 @@ class EngineArgs:
self.tokenizer_pool_type,
self.tokenizer_pool_extra_config,
), self.ray_workers_use_nsight)
scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
self.max_num_seqs,
model_config.max_model_len,
self.use_v2_block_manager,
self.scheduler_delay_factor)
scheduler_config = SchedulerConfig(
self.max_num_batched_tokens,
self.max_num_seqs,
model_config.max_model_len,
self.use_v2_block_manager,
delay_factor=self.scheduler_delay_factor,
enable_chunked_prefill=self.enable_chunked_prefill,
)
lora_config = LoRAConfig(
max_lora_rank=self.max_lora_rank,
max_loras=self.max_loras,
+14 -7
View File
@@ -553,7 +553,10 @@ class LLMEngine:
# Update the scheduled sequence groups with the model outputs.
scheduled_seq_groups = scheduler_outputs.scheduled_seq_groups
for seq_group, outputs in zip(scheduled_seq_groups, output):
for scheduled_seq_group, outputs in zip(scheduled_seq_groups, output):
seq_group = scheduled_seq_group.seq_group
token_chunk_size = scheduled_seq_group.token_chunk_size
seq_group.update_num_computed_tokens(token_chunk_size)
self._process_sequence_group_outputs(seq_group, outputs)
# Free the finished sequence groups.
@@ -561,7 +564,8 @@ class LLMEngine:
# Create the outputs.
request_outputs: List[RequestOutput] = []
for seq_group in scheduled_seq_groups:
for scheduled_seq_group in scheduled_seq_groups:
seq_group = scheduled_seq_group.seq_group
seq_group.maybe_set_first_token_time(now)
request_output = RequestOutput.from_seq_group(seq_group)
request_outputs.append(request_output)
@@ -676,17 +680,20 @@ class LLMEngine:
# Number of Tokens.
if prompt_run:
num_prompt_tokens = sum(
len(seq_group.prompt_token_ids)
for seq_group in scheduler_outputs.scheduled_seq_groups)
len(scheduled_seq_group.seq_group.prompt_token_ids)
for scheduled_seq_group in
scheduler_outputs.scheduled_seq_groups)
num_generation_tokens = sum(
seq_group.num_seqs()
for seq_group in scheduler_outputs.scheduled_seq_groups)
scheduled_seq_group.seq_group.num_seqs()
for scheduled_seq_group in
scheduler_outputs.scheduled_seq_groups)
else:
num_generation_tokens = scheduler_outputs.num_batched_tokens
# Latency Timings.
time_last_iters = []
for seq_group in scheduler_outputs.scheduled_seq_groups:
for scheduled_seq_group in scheduler_outputs.scheduled_seq_groups:
seq_group = scheduled_seq_group.seq_group
# Time since last token.
# (n.b. updates seq_group.metrics.last_token_time)
time_last_iters.append(seq_group.get_last_latency(now))