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

remove _grouped_size_compiled_for_decode_kernels (#3453)

Browse Source
This commit is contained in:
Yineng Zhang
2025-02-10 13:01:21 +08:00
committed by GitHub
parent 52a492a16e
commit 27c4c9cf52
2 changed files with 1 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/start/install.md
View File

@@ -22,7 +22,7 @@ pip install sgl-kernel --force-reinstall --no-deps
pip install -e "python[all]" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer/ pip install -e "python[all]" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer/
``` ```
Note: Please check the [FlashInfer installation doc](https://docs.flashinfer.ai/installation.html) to install the proper version according to your PyTorch and CUDA versions. If you meet with issue like **ImportError: cannot import name `_grouped_size_compiled_for_decode_kernels`**, installing FlashInfer with some older version like 0.1.6 instead of the latest version could solve it. Note: Please check the [FlashInfer installation doc](https://docs.flashinfer.ai/installation.html) to install the proper version according to your PyTorch and CUDA versions.
Note: To AMD ROCm system with Instinct/MI GPUs, do following instead: Note: To AMD ROCm system with Instinct/MI GPUs, do following instead:

15
python/sglang/srt/layers/attention/flashinfer_backend.py
View File

@@ -1077,21 +1077,6 @@ def should_use_tensor_core(
if env_override is not None: if env_override is not None:
return env_override.lower() == "true" return env_override.lower() == "true"
# Try to use _grouped_size_compiled_for_decode_kernels if available
# This is for flashinfer <=0.1.6. Otherwise, there is an accuracy bug
try:
from flashinfer.decode import _grouped_size_compiled_for_decode_kernels
if not _grouped_size_compiled_for_decode_kernels(
num_attention_heads,
num_kv_heads,
):
return True
else:
return False
except (ImportError, AttributeError):
pass
# Calculate GQA group size # Calculate GQA group size
gqa_group_size = num_attention_heads // num_kv_heads gqa_group_size = num_attention_heads // num_kv_heads