EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[misc] Remove useless patch_logits (#4252)

Browse Source 
Torch-npu 2.7.1 has fixed the device check bug. This patch can be
removed now.

- vLLM main:
2918c1b49c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-25 21:25:54 +08:00
committed by GitHub
parent 4864909648
commit 98031653df
3 changed files with 1 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
vllm_ascend/patch/__init__.py
View File

@@ -104,29 +104,7 @@
# Future Plan:
# Remove this patch when vllm merged them.
#
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.v1.sample.sampler.Sampler.gather_logprobs`
# Why:
# We need to patch gather_logprobs to make sure call batched_count_greater_than
# with backend=current_platform.simple_compile_backend
# How：
# Patch gather_logprobs call new batched_count_greater_than
# Related PR (if no, explain why):
# - https://github.com/vllm-project/vllm/pull/21591
# Future Plan:
# Revert it when vLLM merge #21591 and release new version
# ** File: worker/patch_logits.py **
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm._custom_ops.apply_repetition_penalties`
# Why:
# apply_repetition_penalties in vLLM use tensor.is_cuda to check if tensor is on cuda. But the value is always True
# on ascend, thus we need to patch apply_repetition_penalties.
# How：
# Remove the related cuda check in apply_repetition_penalties.
# Related PR (if no, explain why):
# - this is a bug by Ascend only. It can' be fixed in vLLM.
# Future Plan:
# Fix this bug in torch-npu, bump torch-npu version and remove this patch.
# ** File: worker/patch_roberta.py **
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.model_executor.models.roberta.RobertaEmbedding.forward`
# Why: