Repository navigation
Exports on Modal: each XNNPACK and Vulkan window, and the solve, in a Sandbox sized to it - #15
Merged
Merged
Conversation
… Sandbox sized to it The Namespace profile has 15.7 GiB and no swap, which left out every 32k window and most of the models from 1.5B up (finding 40). The GitHub jobs stay on it and send the heavy step to Modal instead (pipeline/remote.py, `python -m pipeline modal-run`): a Sandbox with the memory the window needs, sizing.host_need_bytes times modal.memory_margin, up to the 344064 MiB and 64 cores Modal allows (measured: larger requests are refused), so every window ever built fits again, Qwen3-4B at 32k included. - remote.py builds the image from the pinned requirements (torch from the CPU index, as setup-export does) and the repo's pipeline, uploads inputs to a v2 Volume under a key unique to the run, runs the step with an EXIT trap that commits the Volume with sync on every way out and keeps the step's status (exit 4 included), downloads the outputs, and always stops a live Sandbox and deletes the key. HF_TOKEN goes in as a Secret built from the job's own environment, so nothing is stored on Modal. - export-matrix gives each window its memory_mib; solve-memory sizes the solve at 12 bytes a parameter, as the Blacksmith solves measured. - export-xnnpack.yml: the solve, then one Sandbox per window that exports, smoke-tests through ExecuTorch's C++ runner and gates on the fp32 model's answers before the job publishes. The separate verify job existed for the ARM/x86 split and is gone. export-vulkan.yml sends its export the same way. Both jobs install only the light requirements (setup-modal). - MediaTek and QNN still run on the Namespace profile. - Tested from this workstation on Modal: SmolLM2-135M's Vulkan 2k file came out byte-identical to the published one, and its XNNPACK 2k file exported and passed the smoke test in the same Sandbox; exports were slower than on Blacksmith (finding 41). Codex reviewed it three times; it found the cleanup skipped on a Sandbox timeout, the sync skipped on an explicit exit, the inputs uploaded twice, and the wrong exception caught for a missing Volume path, all fixed and tested. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… Python refuses pip installs Every export workflow's targets job ran `python3 -m pip install` on the runner's Python, which Blacksmith allowed and Namespace does not (PEP 668: externally-managed-environment), so since the runner switch no XNNPACK, Vulkan, MediaTek or QNN workflow could get past its first job. The first workflow run through Modal found it. setup-light gives those jobs Python from versions.env and requirements/dev.txt, which carries the pyyaml and huggingface_hub they used. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
alpharomercoma
marked this pull request as ready for review
October 7, 2026 13:13
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.
XNNPACK and Vulkan exports, and the GPTQ solve, now run on Modal. Each runs in a Sandbox with the memory its window needs (
pipeline/remote.py,python -m pipeline modal-run), up to Modal's 344064 MiB and 64 cores. The GitHub jobs stay on the Namespace profile and only start the Sandbox, wait, download and publish.What changes
setup-light). The Namespace runner's system Python refusespip install(PEP 668), so since the runner switch every export workflow onmainhas stopped at its first job.Verified
Open: per-export time was slower than on Blacksmith (finding 41).
🤖 Generated with Claude Code