Skip to content

vllm bump #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
andre15silva merged 1 commit into from
Dec 31, 2025
Merged

vllm bump #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 7 additions & 4 deletions pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ name = "crrl"
version = "0.1.0"
description = "Finetuning LLMs with RL to improve program repair capabilities"
readme = "README.md"
requires-python = ">=3.11,<3.12"
requires-python = ">=3.11,<3.13"
dependencies = [
# Our fork of trl
"trl @ git+https://github.com/ASSERT-KTH/trl.git@dev",
Expand Down Expand Up @@ -40,9 +40,12 @@ dependencies = [

[project.optional-dependencies]
# vLLM is required, but install is brittle needing torch to be installed first
vllm = ["vllm==0.10.1; platform_system != 'Darwin'"]
flash = ["flash-attn==2.8.3; platform_system != 'Darwin'", "flashinfer-python==0.3.1; platform_system != 'Darwin'"]
# Flash attention packages require GPU at compile time, install separately
vllm = ["vllm==0.11.0; platform_system != 'Darwin'"]
# Flash attention packages - flash-attn needed by transformers for training, flashinfer for vLLM
flash = [
"flash-attn==2.8.3; platform_system != 'Darwin'",
"flashinfer-python==0.5.0; platform_system != 'Darwin'",
]
dev = ["pytest>=8.3.4", "matplotlib>=3.10.0", "scikit-learn>=1.6.1"]
# Different agents
# aider = ["aider-chat @ git+https://github.com/BjarniHaukur/aider.git"]
Expand Down
4 changes: 2 additions & 2 deletions scripts/grpo/large_grpo_lora_train_job.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,9 +55,9 @@ apptainer exec $APPT_COMMON --env CUDA_VISIBLE_DEVICES=2,3,4,5 crrl.sif accelera
grpo.max_prompt_length=$MAX_PROMPT_LENGTH \
grpo.max_completion_length=$MAX_COMPLETION_LENGTH \
grpo.num_generations=4 \
grpo.generation_batch_size=8 \
grpo.generation_batch_size=32 \
grpo.per_device_train_batch_size=1 \
grpo.gradient_accumulation_steps=4 \
grpo.gradient_accumulation_steps=8 \
grpo.optim="adamw_torch" \
"$@" # pass any additional arguments

12 changes: 6 additions & 6 deletions scripts/train_container.def
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
Bootstrap: docker
From: vllm/vllm-openai:v0.10.0
From: vllm/vllm-openai:v0.11.0

%environment
# Keep container largely stateless; runtime env is passed via --env from host
Expand All @@ -15,10 +15,10 @@ From: vllm/vllm-openai:v0.10.0

pip install --upgrade pip
pip install -e .

pip install torch==2.7.1+cu128 --index-url https://download.pytorch.org/whl/cu128

# we pin the version since we need prebuilt binaries to be able to install
pip install --no-build-isolation --prefer-binary flash-attn==2.8.0.post2
pip install --no-build-isolation flashinfer-python
# flash-attn is needed by transformers for attn_implementation="flash_attention_2" during training
# (vLLM only bundles it for inference, not as an importable package)
# Using --no-cache-dir to avoid cross-device link errors in container builds
pip install --no-build-isolation --no-cache-dir flash-attn==2.8.3
pip install --no-build-isolation --no-cache-dir flashinfer-python==0.5.0