[2/N] Chunked prefill data update (#3538)
This commit is contained in:
@@ -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))
|
||||
|
||||
Reference in New Issue
Block a user