feat(quant): pure-quant (no-SC) baselines + 7-bit W/A support - #5
Open
heroarmor wants to merge 1 commit into
Open
feat(quant): pure-quant (no-SC) baselines + 7-bit W/A support#5heroarmor wants to merge 1 commit into
heroarmor wants to merge 1 commit into
Conversation
Add scripts/sbatch_baseline_auto.sb: TIMEOUT-safe auto-resume generation for pure integer-quant baselines (SC never applied — timewise/layerwise unset). Parametrized by <WBITS> <ABITS> <sym|asym>; 2000 imgs, first 200 classes x10 (idx 0-1999), cfg=1.5, seed 0 — same layout as the SC sweeps. FP16 = W16A16. Enable 7-bit: add 7 to --wbits/--abits choices in quant_sc_main.py (was [2,3,4,5,6,8,16]; W7A7 was rejected by argparse). Baseline findings (200-class matched, vs reference): FP16 ~= W8A8 ~= W7A7-asym (near-lossless); asymmetric buys ~1 bit over symmetric; W6 and below (activations <8-bit) collapse — activation quant is the sole driver (W4A8/W6A8 reproduce Q-DiT exactly, weight-only 7-bit == FP16). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
heroarmor
pushed a commit
that referenced
this pull request
Aug 18, 2026
MP-on-DiT: SC halving + global MP budget + pure-quant baselines + kernels #24/#25
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
Adds pure-quant (no-SC) baselines and 7-bit W/A support.
scripts/sbatch_baseline_auto.sb— TIMEOUT-safe auto-resume generation for pure integer-quant baselines (SC never applied:timewise/layerwiseunset). Args:<WBITS> <ABITS> <sym|asym> [CHAIN]. 2000 imgs, first 200 classes ×10 (idx 0-1999), cfg=1.5, seed 0 — identical layout to the SC sweeps.sym=--w_sym --a_sym;asym= neither. FP16 = W16A16.scripts/quant_sc_main.py— add7to--wbits/--abitschoices (was[2,3,4,5,6,8,16]; W7A7 was rejected by argparse).Baseline results (200-class matched idx 0-1999, vs reference)
Takeaways: near-lossless with 8-bit activations (W4A8/W6A8/W8A8 reproduce Q-DiT); asymmetric buys ~1 bit over symmetric; the cliff is activation bit-width, not weight — isolation confirms weight-only 7-bit ≡ FP16 and activation-only 7-bit ≡ the full W7A7 collapse (per-token quant hits outlier channels; no SmoothQuant wired).
Stacked on
feat/mp-global-budget-rawdump(#4).