[Bug] Wheel naming issue

[zhyncs]

ref
#643
#659
#660
#662

@ur4t @xslingcn

Before

flashinfer-0.1.6+cu124torch2.4-cp310-cp310-linux_x86_64.whl

After

flashinfer-0.1.6-cp38-abi3-linux_x86_64.whl

Expected

flashinfer-0.1.6+cu124torch2.4-cp38-abi3-linux_x86_64.whl

[zhyncs]

ref https://github.com/flashinfer-ai/flashinfer/pull/643/files#r1885003167

[zhyncs]

CUDA and Torch version info missing: +cu124torch2.4

[zhyncs]

@ur4t Could you help fix this issue? Thanks.

[xslingcn]

setup.py is taking FLASHINFER_BUILD_VERSION for top priority.

flashinfer/setup.py

Lines 47 to 55 in b1b1fb8

	def get_version():
	build_version = os.environ.get("FLASHINFER_BUILD_VERSION")
	if build_version is not None:
	return build_version
	package_version = (root / "version.txt").read_text().strip()
	local_version = os.environ.get("FLASHINFER_LOCAL_VERSION")
	if local_version is None:
	return package_version
	return f"{package_version}+{local_version}"

It seems we specify this env variable in the nightly building script, while the cuda/torch version is missing:
https://github.com/flashinfer-ai/flashinfer-nightly/blob/f28fab619a21573c9e1ec0dba8bec841c890cfed/.github/workflows/nightly-release.yml#L51-L52

https://github.com/flashinfer-ai/flashinfer-nightly/blob/f28fab619a21573c9e1ec0dba8bec841c890cfed/.github/workflows/nightly-release.yml#L62

[zhyncs]

Yeah I think before these #643 #659 #660 #662 the nightly works

[xslingcn]

Not sure how this was handled before we changed packaging, but one fix I can think of now is:

diff --git a/.github/workflows/nightly-release.yml b/.github/workflows/nightly-release.yml
index 0c919e9..9c6c4ae 100644
--- a/.github/workflows/nightly-release.yml
+++ b/.github/workflows/nightly-release.yml
@@ -48,7 +48,10 @@ jobs:
           sed -i 's/+cu/\.cu/g' scripts/run-ci-build-wheel.sh
           sed -i 's|/ci-cache|/opt/dlami/nvme/flashinfer/github|g' scripts/run-ci-build-wheel.sh
           sed -i '/mkdir -p "\$CONDA_pkgs_dirs" "\$XDG_CACHE_HOME"/d' scripts/run-ci-build-wheel.sh
-          version="$(cat version.txt)"+"$(git rev-parse HEAD | cut -c1-7)"
+          CUDA_VERSION=${{ matrix.cuda }}
+          CUDA_MAJOR="${CUDA_VERSION%.*}"
+          CUDA_MINOR="${CUDA_VERSION#*.}"
+          version="$(cat version.txt)+$(git rev-parse HEAD | cut -c1-7)+cu${CUDA_MAJOR}${CUDA_MINOR}torch${{ matrix.torch }}"
           echo "VERSION=$version" >> $GITHUB_ENV
           echo "version=$version" >> $GITHUB_OUTPUT
         id: get_version

[zhyncs]

There are several ways to ensure nightly compatibility. I want to know the name given when version 0.2.0 was released for the main branch of FlashInfer (not nightly). Also, it would be better if we could be compatible with both the nightly and stable releases.

[xslingcn]

This issue doesn't exist in the stable release workflow, because we don't specify FLASHINFER_BUILD_VERSION so it will use the FLASHINFER_LOCAL_VERSION as passed in:

flashinfer/scripts/run-ci-build-wheel.sh

Line 41 in b1b1fb8

FLASHINFER_ENABLE_AOT=1 FLASHINFER_LOCAL_VERSION="cu${CUDA_MAJOR}${CUDA_MINOR}torch${FLASHINFER_CI_TORCH_VERSION}" python -m build --no-isolation --wheel

[xslingcn]

I want to know the name given when version 0.2.0 was released for the main branch of FlashInfer (not nightly).

It will be something like flashinfer-0.2.0+cu12.4torch2.4-cp38-abi3-linux_x86_64.whl.

[zhyncs]

I see. I'll help with the nightly build issue.

[zhyncs]

fixed with flashinfer-ai/flashinfer-nightly#4