chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Feat]Add support for optional start len of logprobs (#1035)

Browse Source 
Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Yineng Zhang <me@zhyncs.com>
Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
This commit is contained in:
yichuan~
2024-08-18 23:45:41 -07:00
committed by GitHub
parent d8627ed16d
commit b997a18d74
8 changed files with 113 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
python/sglang/srt/layers/logits_processor.py
View File

@@ -55,6 +55,9 @@ class LogitsMetadata:
extend_start_loc: Optional[torch.Tensor] = None
top_logprobs_nums: Optional[List[int]] = None
extend_seq_lens_cpu: List[int] = None
logprob_start_lens_cpu: List[int] = None
@classmethod
def from_input_metadata(cls, input_metadata: InputMetadata):
return cls(
@@ -63,6 +66,8 @@ class LogitsMetadata:
extend_start_loc=input_metadata.extend_start_loc,
return_logprob=input_metadata.return_logprob,
top_logprobs_nums=input_metadata.top_logprobs_nums,
extend_seq_lens_cpu=input_metadata.extend_seq_lens_cpu,
logprob_start_lens_cpu=input_metadata.logprob_start_lens_cpu,
)
@@ -75,12 +80,16 @@ class LogitsProcessor(nn.Module):
)
def _get_normalized_prompt_logprobs(
self, input_token_logprobs, logits_metadata: LogitsMetadata
self,
input_token_logprobs: torch.Tensor,
cum_start_len0: torch.Tensor,
cum_start_len1: torch.Tensor,
logits_metadata: LogitsMetadata,
):
logprobs_cumsum = torch.cumsum(input_token_logprobs, dim=0, dtype=torch.float32)
start = logits_metadata.extend_start_loc.clone()
end = start + logits_metadata.extend_seq_lens - 2
start = logits_metadata.extend_start_loc.clone() - cum_start_len0
end = start + logits_metadata.extend_seq_lens - 2 - cum_start_len1
start.clamp_(min=0, max=input_token_logprobs.shape[0] - 1)
end.clamp_(min=0, max=input_token_logprobs.shape[0] - 1)
sum_logp = (
@@ -93,7 +102,7 @@ class LogitsProcessor(nn.Module):
return normalized_prompt_logprobs
@staticmethod
def get_top_logprobs(all_logprobs, logits_metadata: LogitsMetadata):
def get_top_logprobs(all_logprobs: torch.Tensor, logits_metadata: LogitsMetadata):
if logits_metadata.forward_mode == ForwardMode.DECODE:
output_top_logprobs = []
max_k = max(logits_metadata.top_logprobs_nums)
@@ -107,7 +116,7 @@ class LogitsProcessor(nn.Module):
# TODO: vectorize the code below
input_top_logprobs, output_top_logprobs = [], []
pt = 0
extend_seq_lens_cpu = logits_metadata.extend_seq_lens.tolist()
extend_seq_lens_cpu = logits_metadata.extend_seq_lens_cpu
max_k = max(logits_metadata.top_logprobs_nums)
ret = all_logprobs.topk(max_k, dim=1)
@@ -115,26 +124,30 @@ class LogitsProcessor(nn.Module):
indices = ret.indices.tolist()
for i, extend_seq_len in enumerate(extend_seq_lens_cpu):
start_len = logits_metadata.logprob_start_lens_cpu[i]
pruned_len = extend_seq_len - start_len
if extend_seq_len == 0:
input_top_logprobs.append([])
output_top_logprobs.append([])
continue
k = logits_metadata.top_logprobs_nums[i]
input_top_logprobs.append(
[
list(zip(values[pt + j][:k], indices[pt + j][:k]))
for j in range(extend_seq_len - 1)
for j in range(pruned_len - 1)
]
)
output_top_logprobs.append(
list(
zip(
values[pt + extend_seq_len - 1][:k],
indices[pt + extend_seq_len - 1][:k],
values[pt + pruned_len - 1][:k],
indices[pt + pruned_len - 1][:k],
)
)
)
pt += extend_seq_len
pt += pruned_len
return input_top_logprobs, output_top_logprobs
@@ -205,7 +218,23 @@ class LogitsProcessor(nn.Module):
output_top_logprobs=output_top_logprobs,
)
else:
all_logits = torch.matmul(hidden_states, weight.T)
pt, states, pruned_input_ids = 0, [], []
for i, extend_len in enumerate(logits_metadata.extend_seq_lens_cpu):
start_len = logits_metadata.logprob_start_lens_cpu[i]
states.append(hidden_states[pt + start_len : pt + extend_len])
pruned_input_ids.append(input_ids[pt + start_len : pt + extend_len])
pt += extend_len
states = torch.cat(states, dim=0)
pruned_input_ids = torch.cat(pruned_input_ids, dim=0)
cum_start_len1 = torch.tensor(
logits_metadata.logprob_start_lens_cpu, device="cuda"
).cumsum(0)
cum_start_len0 = torch.zeros_like(cum_start_len1)
cum_start_len0[1:] = cum_start_len1[:-1]
all_logits = torch.matmul(states, weight.T)
if self.do_tensor_parallel_all_gather:
all_logits = tensor_model_parallel_all_gather(all_logits)
all_logits = all_logits[:, : self.config.vocab_size].float()
@@ -230,19 +259,25 @@ class LogitsProcessor(nn.Module):
else:
input_top_logprobs = output_top_logprobs = None
last_logprobs = all_logprobs[last_index]
last_logprobs = all_logprobs[last_index - cum_start_len1]
# Compute the logprobs and normalized logprobs for the prefill tokens.
# Note that we pad a zero at the end of each sequence for easy computation.
input_token_logprobs = all_logprobs[
torch.arange(all_logprobs.shape[0], device="cuda"),
torch.cat([input_ids[1:], torch.tensor([0], device="cuda")]),
torch.cat([pruned_input_ids[1:], torch.tensor([0], device="cuda")]),
]
normalized_prompt_logprobs = self._get_normalized_prompt_logprobs(
input_token_logprobs, logits_metadata
input_token_logprobs,
cum_start_len0,
cum_start_len1,
logits_metadata,
)
# Remove the last token logprob for the prefill tokens.
input_token_logprobs = input_token_logprobs[:-1]
return LogitProcessorOutput(
next_token_logits=last_logits,
next_token_logprobs=last_logprobs,