docs: fix manual-installation.md missing batchgen_kernels build step - #278
Merged
Conversation
…ersions Verified in a fresh Ubuntu 22.04 + CUDA 12.8 container (no SSH keys): - the three 'git clone git@github.com:...' commands fail on a fresh server; switch to https:// like INSTALL.md and the Dockerfile - unpinned FlashMLA HEAD fails to build with nvcc 12.8 (SM100 kernels need 12.9+); pin commit 1408756a + FLASH_MLA_DISABLE_SM100=1, matching scripts/install_deps.sh - unpinned DeepGEMM HEAD (54e22612) fails to build; pin v2.1.1.post3, which builds fine - soften the FORCE_BUILD note: the verbatim no-env build completed in 29 min in the same container (the wheel-download attempt fell back), so 'hangs' is situational, not guaranteed refs #277
18 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
docs/manual-installation.mdhad drifted from the real install flow. Four fixes:batchgen_kernelsbuild step.batchgen_kernelsis a separate package (batchgen_kernels/setup.py, 22 AOT CUDA extensions) that the top-levelpip install .does not build (mainsetup.pyonly packagesbatchgen/batchgen.*). The guide previously skipped it entirely, producing an install with all compiled extensions missing → runtime import failures. Added a dedicated step (matchinginstall_deps.sh'sinstall_batchgen_kernels), with theTORCH_CUDA_ARCH_LIST=9.0a(H20) /BUILD_ARCH=sm100(Blackwell) notes.pip install -e .→ non-editablepip install .. Aligns with INSTALL.md, which requires non-editable for ray/production and warns the source dir shadows installed packages. Added a one-line caveat + link.FLASH_ATTENTION_FORCE_BUILD=TRUEto the FA3 hopper build (against torch 2.9 it otherwise hangs downloading a non-existent prebuilt wheel — same asinstall_deps.sh:187/ INSTALL.md troubleshooting).All cross-reference anchors and the
BUILD_ARCHknob were verified against the code. Thegit clone --recursivefor DeepGEMM is correct (DeepGEMM genuinely has submodules) and left unchanged.Motivation
Following the manual guide verbatim currently yields a non-functional install. Resolves #277.
close #277
Type of Change
modelkernelcorefixinfradocs— documentation onlyFile changes
docs/manual-installation.mdFLASH_ATTENTION_FORCE_BUILD=TRUE; verification pointerChecklist
PR Merge Policy Contract — pre-merge checklist
git diff --stat origin/mainreviewed; every change traces to the task (§3.1).docs); onlydocs/**touched (§2.5).print(), or artifacts added (§1).Co-Authored-By(§4; body optional fordocs).INSTALL.md#...) andBUILD_ARCHknob verified against the repo.