# xv01-train: training image for xevyo-01 (LoRA on Qwen/Qwen3.5-9B). # # docker build -t xv01-train:latest /data/xevyo-01/code/train # # Base: official PyTorch image, torch 2.10 (same torch as the vLLM 0.17 image we # serve with), CUDA 12.8, devel variant so causal-conv1d can compile for sm90 if # its prebuilt wheel is ever unavailable. FROM pytorch/pytorch:2.10.0-cuda12.8-cudnn9-devel # PIP_BREAK_SYSTEM_PACKAGES: this base image installs torch into the system # Python 3.12 (/usr/local/lib/python3.12/dist-packages), which Ubuntu marks as # externally managed (PEP 668). Overriding that is safe inside a single-purpose # image and keeps the base image's torch build. ENV PIP_NO_CACHE_DIR=1 \ PIP_BREAK_SYSTEM_PACKAGES=1 \ PIP_DISABLE_PIP_VERSION_CHECK=1 \ PYTHONUNBUFFERED=1 \ PYTHONDONTWRITEBYTECODE=1 \ TORCH_CUDA_ARCH_LIST="9.0" \ MAX_JOBS=16 \ HF_HOME=/data/xevyo-01/hf \ HF_HUB_OFFLINE=1 \ TOKENIZERS_PARALLELISM=false # Freeze torch/triton exactly as the base image ships them, so no requirement # below can pull a different torch build. RUN python -c "import importlib.metadata as m; print('\n'.join(f'{p}=={m.version(p)}' for p in ('torch','triton')))" > /opt/constraints.txt \ && cat /opt/constraints.txt COPY requirements.txt /opt/xv01/requirements.txt RUN pip install -c /opt/constraints.txt -r /opt/xv01/requirements.txt # causal-conv1d's setup.py downloads the matching prebuilt wheel # (cu12 / torch 2.10 / cxx11abi) and falls back to compiling; it needs torch # importable at build time, hence --no-build-isolation. RUN pip install -c /opt/constraints.txt --no-build-isolation causal-conv1d==1.7.0 # Triton 3.7.1, on purpose newer than the 3.6.0 that torch 2.10 pins. # flash-linear-attention refuses to run the Gated DeltaNet backward kernel # (chunk_bwd_dqkwg) on Hopper GPUs (H100) with Triton >= 3.4.0 and < 3.7.1, # because those versions compile it to wrong gradients (fla issue #640). The first # smoke run on 2026-11-13 died on exactly that error. pip prints a "torch requires # triton==3.6.0" warning here; that pin only matters for torch.compile, which the # training code does not use. Installed last and without the constraints file. RUN pip install triton==3.7.1 # Fail the build if the stack does not import, and print the versions. RUN python -c "import torch, transformers, peft, trl, accelerate, datasets, fla, causal_conv1d; \ print('torch', torch.__version__, 'cuda', torch.version.cuda); \ print('transformers', transformers.__version__, 'peft', peft.__version__, 'trl', trl.__version__); \ print('accelerate', accelerate.__version__, 'datasets', datasets.__version__, 'fla', fla.__version__); \ import triton; print('triton', triton.__version__); assert tuple(int(x) for x in triton.__version__.split('.')[:3]) >= (3, 7, 1); \ from transformers import Qwen3_5ForConditionalGeneration, Qwen3_5ForCausalLM; print('qwen3_5 classes ok')" \ && pip freeze > /opt/xv01/pip-freeze.txt WORKDIR /data/xevyo-01/code