EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

hotfix: revert sampler CUDA Graph (#1242)

Browse Source
This commit is contained in:
Yineng Zhang
2024-08-28 21:16:47 +10:00
committed by GitHub
parent 184ae1c683
commit f25f4dfde5
33 changed files with 119 additions and 348 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/bench_latency.py
View File

@@ -200,14 +200,16 @@ def extend(reqs, model_runner):
tree_cache=None,
)
batch.prepare_for_extend(model_runner.model_config.vocab_size)
sample_output, logits_output = model_runner.forward(batch, ForwardMode.EXTEND)
return sample_output.batch_next_token_ids, logits_output.next_token_logits, batch
output = model_runner.forward(batch, ForwardMode.EXTEND)
next_token_ids = batch.sample(output.next_token_logits)
return next_token_ids, output.next_token_logits, batch
def decode(input_token_ids, batch, model_runner):
batch.prepare_for_decode(input_token_ids.cpu().numpy())
sample_output, logits_output = model_runner.forward(batch, ForwardMode.DECODE)
return sample_output.batch_next_token_ids, logits_output.next_token_logits
output = model_runner.forward(batch, ForwardMode.DECODE)
next_token_ids = batch.sample(output.next_token_logits)
return next_token_ids, output.next_token_logits
@torch.inference_mode()