Delay GPU->CPU sync in sampling (#1337)
This commit is contained in:
@@ -109,29 +109,8 @@ def _prune_hidden_states(
|
||||
hidden_states: torch.Tensor,
|
||||
input_metadata: InputMetadata,
|
||||
) -> torch.Tensor:
|
||||
selected_token_indices: List[int] = []
|
||||
start_idx = 0
|
||||
for i, seq_group in enumerate(input_metadata.seq_groups):
|
||||
seq_ids, sampling_params = seq_group
|
||||
if i < input_metadata.num_prompts:
|
||||
assert len(seq_ids) == 1, "Prompt input should have only one seq."
|
||||
prompt_len = input_metadata.prompt_lens[i]
|
||||
if sampling_params.prompt_logprobs is not None:
|
||||
selected_token_indices.extend(
|
||||
range(start_idx, start_idx + prompt_len - 1))
|
||||
selected_token_indices.append(start_idx + prompt_len - 1)
|
||||
start_idx += input_metadata.max_prompt_len
|
||||
else:
|
||||
num_seqs = len(seq_ids)
|
||||
selected_token_indices.extend(
|
||||
range(start_idx, start_idx + num_seqs))
|
||||
start_idx += num_seqs
|
||||
|
||||
selected_token_indices = torch.tensor(selected_token_indices,
|
||||
dtype=torch.long,
|
||||
device=hidden_states.device)
|
||||
hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
|
||||
return hidden_states.index_select(0, selected_token_indices)
|
||||
return hidden_states.index_select(0, input_metadata.selected_token_indices)
|
||||
|
||||
|
||||
def _get_penalties(
|
||||
@@ -426,21 +405,11 @@ def _sample(
|
||||
input_metadata: InputMetadata,
|
||||
) -> List[Tuple[List[int], List[int]]]:
|
||||
categorized_seq_group_ids = {t: [] for t in SamplingType}
|
||||
categorized_sample_indices = {t: [] for t in SamplingType}
|
||||
start_idx = 0
|
||||
categorized_sample_indices = input_metadata.categorized_sample_indices
|
||||
for i, seq_group in enumerate(input_metadata.seq_groups):
|
||||
seq_ids, sampling_params = seq_group
|
||||
_, sampling_params = seq_group
|
||||
sampling_type = sampling_params.sampling_type
|
||||
if (i < input_metadata.num_prompts
|
||||
and sampling_params.prompt_logprobs is not None):
|
||||
# NOTE: prompt token positions do not need sample, skip
|
||||
prompt_len = input_metadata.prompt_lens[i]
|
||||
start_idx += prompt_len - 1
|
||||
categorized_seq_group_ids[sampling_type].append(i)
|
||||
num_seqs = len(seq_ids)
|
||||
categorized_sample_indices[sampling_type].extend(
|
||||
range(start_idx, start_idx + num_seqs))
|
||||
start_idx += num_seqs
|
||||
|
||||
sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
|
||||
for sampling_type in SamplingType:
|
||||
|
||||
Reference in New Issue
Block a user