[fix(docker)]: build fine-tune env from source instead of the unpublished release wheel

[devangpratap]

What does this PR do?

Fixes #1762.

The docker/Dockerfile sft build downloads a prebuilt ktransformers wheel:

https://github.com/kvcache-ai/ktransformers/releases/download/v0.5.3/ktransformers-0.5.3+cu128torch28fancy-cp312-cp312-linux_x86_64.whl

No ktransformers release (v0.5.0 through v0.6.2) publishes wheel assets, so that URL returns 404 and curl -fsSL aborts the build.

On current main the top-level ktransformers package is a thin metapackage (kt-kernel plus [sft] extras), and the repo is already cloned into the image. This installs it from that source instead of the missing wheel, and aligns the fine-tune env with the dependencies current kt-kernel pins.

Changes

• Remove the KTRANSFORMERS_VERSION / KTRANSFORMERS_WHEEL ARGs and the dead wheel download.
• Build the local kt-kernel from source in the fine-tune env (same as the serve env), then install ktransformers[sft] from the cloned /workspace/ktransformers. This uses the checked-out C++ kernels and avoids pip falling back to a PyPI kt-kernel that may not exist for an unreleased version.
• Bump the fine-tune env to torch==2.9.1, which is what kt-kernel (pulled in by [sft]) pins.
• Switch FLASH_ATTN_WHEEL to the matching torch2.9 build from the same v2.8.3 release.
• Add editables to the build tools, required by the editable --no-build-isolation install of LLaMA-Factory.
• Drop the removed wheel from the cleanup step.

What was tested

Replicated the fine-tune env dependency resolution in a nvidia/cuda:12.8.1-cudnn-devel-ubuntu24.04 container (CUDA 12.8, Python 3.12), running the same commands in the same order as the Dockerfile fine-tune stage:

• torch==2.9.1 (cu128), then LLaMA-Factory[torch,metrics], then ktransformers[sft], then the flash_attn torch2.9 wheel.
• All install with exit code 0. torch stays at 2.9.1+cu128 after every step (no silent up/downgrade).
• kt_kernel, llamafactory, transformers (transformers-kt 5.6.0), and accelerate (accelerate-kt 1.14.0) import.
• pip check reports no broken requirements.

This confirms the torch/flash_attn/editables alignment and that ktransformers[sft] plus LLaMA-Factory coexist under torch 2.9.1. In that run kt-kernel came from PyPI as a stand-in for dependency resolution.

What was not tested

• A full docker build of this Dockerfile was not run. The base image (docker.1ms.run) and the tsinghua conda/pip mirrors are not reachable from the test machine, and the single framework stage also compiles DeepEP, kt-kernel (all CPU variants), and sglang for the serve env, which this PR does not touch.
• The local kt-kernel source build in the fine-tune env (CPUINFER_BUILD_ALL_VARIANTS=1 ./install.sh build) was not run here. It mirrors the existing serve-env build step exactly, but it is a heavy C++/CUDA compile that was not executed in this verification.
• The editables addition depends on the hatchling version resolved at build time. It was required to reproduce a clean editable install here; it is harmless if the mirror-resolved hatchling would not have needed it.

Before submitting

• Read the contributor guideline.
• Did you write any new necessary tests? No automated test added; a Dockerfile change is verified by reproducing the install sequence as described above.

[gemini-code-assist]

Code Review

This pull request updates the Dockerfile to install ktransformers from the cloned source tree instead of downloading a prebuilt wheel, and upgrades PyTorch to version 2.9.1 along with the corresponding flash_attn wheel. A critical issue was identified where installing ktransformers[sft] directly without building the local kt-kernel package first will cause pip to download kt-kernel from PyPI, ignoring local C++ kernel modifications. It is recommended to build and install the local kt-kernel in the fine-tune environment prior to installing ktransformers.

[gemini-code-assist]

Installing ktransformers[sft] directly from /workspace/ktransformers[sft] without first building and installing the local kt-kernel package in the fine-tune environment will cause pip to resolve and download kt-kernel from PyPI (or the configured mirror). This means any local modifications to the C++ kernels in kt-kernel/ will be completely ignored during fine-tuning, and it could also fail if the matching version of kt-kernel is not yet published.

To ensure consistency and that local changes are correctly built and used, we should build and install the local kt-kernel in the fine-tune environment using the same install.sh script, just as is done for the serve environment.

RUN --mount=type=cache,target=/root/.cache/pip \
    if [ "$FUNCTIONALITY" = "sft" ]; then \
        . /opt/miniconda3/etc/profile.d/conda.sh && conda activate fine-tune \
        && cd /workspace/ktransformers/kt-kernel \
        && CPUINFER_BUILD_ALL_VARIANTS=1 ./install.sh build \
        && pip install "/workspace/ktransformers[sft]"; \
    fi

[devangpratap]

@jdai0 sorry for the ping, this is ready for review, just needs the run-ci label. Thanks

[jdai0]

Thanks, I've added the run-ci label.