FEP(sig-framework): FlagOS-Compressor model quantization framework - #91
Open
wbavon wants to merge 7 commits into
Open
FEP(sig-framework): FlagOS-Compressor model quantization framework#91wbavon wants to merge 7 commits into
wbavon wants to merge 7 commits into
Conversation
|
|
wbavon
added
FEP
Verified every claim against the release/v0.1.0 branch (eecf2fa) rather than main: - INT8, per-channel INT8, and dynamic-token W8A8 (PRs #3/#5) are in the release branch, so they move from Non-Goals into G2 with their actual storage formats and defaults. - Calibrated GPTQ/AWQ/AutoRound (PR #6) landed on main after the branch was cut; recorded as out of scope for 2.2. - Backends are cpu and cuda only; other chips are covered by generating on a CPU/CUDA host and validating the load on target. - G4 states the fused-3D-expert limit: only the Qwen3.5-MoE layout is registered, other axis orders are refused. - Test Plan now carries the exact commands and expected results the template requires, including the guard/negative cases. - Packaging answers the wheel question; PRs #4/#5 added to Related PRs.
- Drop the (Required)/(Optional) template markers; no merged FEP keeps
them.
- Remove the FlagCompressor rename note, which has no bearing on the
proposal.
- Cut editorializing ("deliberate", "conservative baseline", "not to be
bypassed", "machine-readable", "correct behavior") and redundant
clauses, keeping the technical claims unchanged.
wbavon
force-pushed
the
fep-flagos-compressor
branch
from
August 25, 2026 06:11
9b727e6 to
0e8d8e3
Compare
Collaborator
Author
|
recheck |
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
FEP for the first release (v0.1.0) of FlagOS-Compressor, the FlagOS weight-format conversion and quantization tool, in the FlagOS 2.2 cycle.
Scope is pinned to the release branch
release/v0.1.0(cut ateecf2fa, taggedv0.1.0-rc1), notmain:pack-quantized), plus dynamic per-token W8A8 (int-quantized).linear,attention,mlp,moe,moe.routed,moe.shared), regex select/exclude, and versioned YAML recipes.inspect,--dry-run,validate; CPU and CUDA backends.Notable scope corrections made by checking the release branch rather than
main:mainafter the branch was cut and is out of scope for 2.2.--backendacceptscpuandcudaonly; other chips are covered by generating on a CPU/CUDA host and validating the load on target.New module and repository, so this FEP doubles as the module admission proposal.
Status is
Implementable: implementation is complete on the release branch (PRs #1–#5 merged,v0.1.0-rc1tagged 2026-08-24) and the Test Plan carries executable commands with expected results. Owner: @rdzhu225. Acceptance runs during the 2.2 testing window flip it toImplemented. Open TODOs carried in doc comments, to be settled during acceptance: numerical tolerance for the dequant/quant checks, host-memory sizing guidance, dependency pins / wheel publication, and the serving stack + datasets for the load tests. Feature Freeze for 2.2 is 2026-08-31.