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

Minor follow-up fixes for the logprob refactor (#2670)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-12-30 05:42:08 -08:00
committed by GitHub
parent c5210dfa38
commit 21ec66e59e
5 changed files with 11 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/logits_processor.py
View File

@@ -35,21 +35,21 @@ from sglang.srt.model_executor.forward_batch_info import (
@dataclasses.dataclass
class LogitsProcessorOutput:
## First part. This part will be returned by python/sglang/srt/layers/logits_processor.py::LogitsProcessor.
## Part 1: This part will be assigned in python/sglang/srt/layers/logits_processor.py::LogitsProcessor
# The logits of the next tokens. shape: [#seq, vocab_size]
next_token_logits: torch.Tensor
# Used by speculative decoding (EAGLE)
# The last hidden layers
hidden_states: Optional[torch.Tensor] = None
## Second part. This part will be returned by python/sglang/srt/layers/sampler.py::Sampler.
## Part 2: This part will be assigned in python/sglang/srt/layers/sampler.py::Sampler
# The logprobs of the next tokens. shape: [#seq]
next_token_logprobs: Optional[torch.Tensor] = None
# The logprobs and ids of the top-k tokens in output positions. shape: [#seq, k]
next_token_top_logprobs_val: Optional[List] = None
next_token_top_logprobs_idx: Optional[List] = None
## Third part. This part will be returned by python/sglang/srt/layers/logits_processor.py::LogitsProcessor. Prefill-only.
## Part 3: Prefill-only. This part will be assigned in python/sglang/srt/layers/logits_processor.py::LogitsProcessor
# The normlaized logprobs of prompts. shape: [#seq]
normalized_prompt_logprobs: torch.Tensor = None
# The logprobs of input tokens. shape: [#token]

4
python/sglang/srt/layers/sampler.py
View File

@@ -56,7 +56,9 @@ class Sampler(nn.Module):
if global_server_args_dict["sampling_backend"] == "flashinfer":
if return_logprob:
# NOTE: the top_p_renorm_prob from flashinfer has numerical problems
# NOTE: the top_p_renorm_prob from flashinfer has numerical problems,
# https://github.com/flashinfer-ai/flashinfer/issues/708
# so we use the torch implementation.
logprobs = torch.log(
top_p_normalize_probs_torch(probs, sampling_info.top_ps)
)

7
python/sglang/srt/model_executor/model_runner.py
View File

@@ -36,7 +36,7 @@ from sglang.srt.layers.attention.flashinfer_backend import FlashInferAttnBackend
from sglang.srt.layers.attention.torch_native_backend import TorchNativeAttnBackend
from sglang.srt.layers.attention.triton_backend import TritonAttnBackend
from sglang.srt.layers.logits_processor import LogitsProcessorOutput
from sglang.srt.layers.sampler import Sampler, get_top_logprobs
from sglang.srt.layers.sampler import Sampler
from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
from sglang.srt.lora.lora_manager import LoRAManager
from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -191,10 +191,9 @@ class ModelRunner:
torch.get_device_module(self.device).set_device(self.gpu_id)
if self.device == "cuda":
backend = "nccl"
# TODO(liangan1):Just use gloo to bypass the initilization fail
# Need to use xccl for xpu backend in the future
elif self.device == "xpu":
# TODO(liangan1):Just use gloo to bypass the initilization fail
# Need to use xccl for xpu backend in the future
backend = "gloo"
elif self.device == "hpu":
backend = "hccl"

4
python/sglang/srt/sampling/sampling_batch_info.py
View File

@@ -244,7 +244,7 @@ class SamplingBatchInfo:
# repetition
if self.scaling_penalties is not None:
logits = torch.where(
logits[:] = torch.where(
logits > 0,
logits / self.scaling_penalties,
logits * self.scaling_penalties,
@@ -253,5 +253,3 @@ class SamplingBatchInfo:
# Apply regex vocab_mask
if self.vocab_mask is not None:
self.apply_mask(logits=logits, vocab_mask=self.vocab_mask)
return logits