Skip to content

chore(scripts): restore PR #65 (Q8_0 default + --target flag) lost in ht rewrite#88

Closed
marksverdhei wants to merge 1 commit into
htfrom
chore/restore-bench-dflash-q8
Closed

chore(scripts): restore PR #65 (Q8_0 default + --target flag) lost in ht rewrite#88
marksverdhei wants to merge 1 commit into
htfrom
chore/restore-bench-dflash-q8

Conversation

@marksverdhei
Copy link
Copy Markdown

Summary

Epoch #86 task 3 (bench-dflash.sh drift check) surfaced this: PR #65 ("scripts(dflash): switch default bench target to Q8_0 + --target flag") was merged 2026-06-04 19:39 (merge commit `09b2124f`), but its content is missing from the current `origin/ht` tip. The script on ht still has the pre-#65 state — `TARGET=...gemma-4-31B-it-Q4_K_M.gguf` hardcoded, no `--target` flag, no `DFLASH_BENCH_TARGET` env var.

Likely cause: the 2026-06-04 ht history rewrite happened around the same time and didn't carry PR #65's diff through the 8-commit consolidation. (Spot-checked PRs #52, #54, #55 from the same window — those are all still present under different commit hashes, so this is genuinely the one that fell through.)

Fix

Cherry-picked `2079a61de` ("scripts(dflash): switch default bench target to Q8_0 + --target flag") directly onto ht. Same diff Markus originally approved; no other changes.

Test plan

  • `git cherry-pick 2079a61` clean (no conflicts).
  • Post-pick `bench-dflash.sh` has Q8_0 default, `--target PATH` flag, `DFLASH_BENCH_TARGET` env var, `DFLASH_BENCH_DRAFTER_DIR` env var, and the VRAM-math comment block.
  • (post-merge) re-run bench-dflash on a Q8 target to confirm the lower default still works.

Audit follow-up

Suggest tagging `ht-pre-rewrite-2026-06-04` (already exists per memory) as the canonical reference for "what content was present pre-rewrite" — so future rewrites can sanity-check against this list. Filing as a low-priority observation rather than a PR.

🤖 Generated with Claude Code

Per Markus 2026-06-04: DFlash quality measurement should use a Q8_0
target rather than Q4_K_M, since Q4_K_M introduces enough target-side
quantization noise to confound DFlash's own accept-rate signal. Q8_0
fits in 38 GB total, well within titan A100 80 GB.

* Default `TARGET` is now `gemma-4-31B-it-Q8_0.gguf`. Override via
  `--target PATH` or `DFLASH_BENCH_TARGET` env var.
* Also added `DFLASH_BENCH_DRAFTER_DIR` env var for consistency.
* Comment block documents VRAM math for Q4_K_M / Q8_0 / BF16 targets
  so future runs can pick the right card.
@marksverdhei
Copy link
Copy Markdown
Author

Retracting — false positive. My local origin/ht was stale at the time I ran the ancestry check; after a fresh fetch + pull, origin/ht IS at 09b2124fb which includes the original PR #65 merge commit. The script already has the Q8_0 default and --target flag.

This is exactly the failure mode feedback_local_vs_origin_branch_check was written to prevent — I checked against local refs without fetching first. Updating the memory note to reflect that the cherry-pick was unnecessary.

Apologies for the noise; deleting the branch.

@marksverdhei marksverdhei deleted the chore/restore-bench-dflash-q8 branch June 6, 2026 00:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant