📋 FlyDSL upstream tracker — jhinpan issues & PRs

[jhinpan]

Living tracker for all FlyDSL (ROCm/FlyDSL) issues filed and PRs opened by @jhinpan. Last updated 2026-06-17.

Tally: 11 issues, 22 PRs (= 18 distinct PR work items; 4 PRs are closed->reopened duplicates, see ♻️).

Recent tracker corrections since 2026-06-11: PR #639 and PR #675 are now merged; PR #670 is closed unmerged because merged PR #683 absorbed the relevant flash-attn work; PR #637 is closed unmerged; issue #614 is closed.

Live refresh 2026-06-17: PR #685 is no longer a draft — now ready for review (retitled to batch-aware dense seq_len routing — DUALWAVE_SWP vs generic; mergeable, review required, no CI checks reported yet). No new jhinpan-authored issues/PRs since the last update; all other PR/issue states below are unchanged and re-verified against GitHub.

---

🟢 Open PRs / review-CI state

• FMHA gfx950: batch-aware dense seq_len routing (DUALWAVE_SWP vs generic) (ready for review) — PR #685 (no issue) · un-drafted 2026-06-17; mergeable, review required (branch-protection blocked), no CI checks reported yet
• Complete BasisAttr support in IntTupleBuilder — PR #638 → issue #574 · GitHub marks it conflicting / dirty against main; last CI run is stale (mixed pass/fail across reruns, untouched since 2026-06-11) — needs a rebase before merge review can move
• Layout-algebra inference diagnostics / verifiers (draft) — PR #648 → issue #583 · review required, no checks reported yet

✅ Merged

• rmsnorm: known_block_size on large-M small-N path — PR #639 → fixed issue #614 · merged 2026-06-16; #614 closed 2026-06-16
• CI performance dashboard at /ci-dashboard — PR #675 (no issue) · merged 2026-06-16
• Onboarding notebooks (2/n): layout algebra — PR #665 → issue #573 · merged 2026-06-11; #573 kept open for remaining notebooks
• Onboarding notebooks (1/n): expr foundation — PR #635 → issue #573 · merged 2026-06-06; 1/n — #573 kept open for remaining notebooks
• run_benchmark: report base norm bandwidth, not the last variant — PR #654 → fixed issue #655 · merged 2026-06-06; CI layernorm 1.69 -> 5.6 TB/s
• docs: refresh CLAUDE.md (2-month changes + expr-neutral / helper-placement conventions) — PR #659 (no issue) · merged 2026-06-06
• softmax: enable vectorized fast path — PR #650 → fixed issue #627 · merged 2026-06-04
• moe_blockscale: fix broken e2e test — PR #643 → fixed issue #642 · merged 2026-06-04
• vec_add wider vector copy chunks — PR #564 → issue #563 (issue by others) · merged 2026-06-02
• CI: allow PR authors to request failed CI reruns — PR #602 (no issue) · merged 2026-06-02
• docs: sync compile pipeline references — PR #572 (no issue) · merged 2026-05-27

❌ Closed unmerged / inactive PRs

• flash_attn gfx950: dwordx4 O stores + flash-decoding split-K (draft) — PR #670 (no issue) · closed 2026-06-16 unmerged because merged #683 absorbed this line of work, including pieces from the draft such as num_kv_splits plumbing and the gfx950 O-store work; CI was green before close, but the standalone branch is now conflicting / dirty
• Fix stale JIT dependency caches — PR #637 → issue #453 (issue by others) · closed 2026-06-16 unmerged after maintainer note that another PR already fixed it; #453 itself is still open
• source_loc: finer ROCprof ATT source mapping — PR #593 → issue #587 · closed 2026-06-02 unmerged; #587 was closed via upstream PR #586
• Include static tensor layouts in JIT cache keys — PR #562 → issue #317 (issue by others) · closed 2026-05-28 unmerged; #317 later closed

🐛 Issues — status

Issue	State	Addressed by
#655 run_benchmark mislabels layernorm bandwidth	closed	#654 ✅
#653 device printf invisible in Jupyter / piped stdout	open	— (no PR yet)
#642 moe_blockscale e2e harness bug	closed	#643 ✅
#627 softmax fast path dead-coded off	closed	#650 ✅
#614 rmsnorm large-M small-N crash	closed	#639 ✅
#612 Discussion: make autotune usable	open	— (discussion)
#587 ATT source-location granularity	closed	upstream #586 ✅; jhinpan #593 ❌ closed unmerged
#585 const-folded fp8 cast rounding	closed	— (working-as-intended)
#583 layout-algebra op verification	open	#648 🟢 draft
#574 Complete BasisAttr support	open	#638 🟡 open; conflicts with main (CI stale), needs rebase
#573 Onboarding Jupyter notebook	open	#635 ✅ (1/n merged), #665 ✅ (2/n merged); open for remaining notebooks

♻️ Superseded PR pairs (closed -> reopened, same work)

The old head repo was deleted during a fork rename, so GitHub lost the head association and these were recreated on the same branch:

• #615 -> #639 ✅ merged (issue #614 closed)
• #605 -> #638 🟡 open; conflicts with main (issue #574 open)
• #584 -> #635 ✅ merged (issue #573 still open for remaining notebooks)
• #565 -> #637 ❌ closed unmerged after maintainer said another PR already fixed it (issue #453 still open)

📌 Follow-ups to decide

• #638 / #574 — rebase or refresh branch; GitHub marks the PR conflicting / dirty and its CI is stale.
• #685 — flash-attn/FMHA routing PR is now ready for review (un-drafted 2026-06-17). Next: nudge a reviewer / confirm CI starts on the self-hosted GPU runners.
• #648 / #583 — draft verifier PR with no checks yet; decide whether to continue or rescope.
• #670 — closed flash_attn perf draft; no standalone revival needed for the absorbed work because merged #683 already picked up the relevant pieces. Revisit only if there is a post-#683 gap.
• #637 / #453 — PR closed as already fixed elsewhere, but the linked issue remains open; decide whether to verify and close/ask upstream.
• #653 (device printf invisible in Jupyter) — no PR yet; relates to onboarding-notebook UX. Decide whether to drive a fix (libc stdout buffering).
• #612 (autotune) — discussion; decide whether to drive to a PR.
• #573 (onboarding notebook series) — 1/n and 2/n are merged; continue with remaining notebooks.
• #585 (const-fold fp8 cast) — resolved: closed working-as-intended (apply fp8/bf16 rounding on runtime values, not a compile-time constant round-trip)