Repository navigation
[Other] [P1][UPSTREAM][RTX5080] Upstream ancestry and post-v0.1.44.1 sync #23
Description
Finding — 2026-10-17
GitHub compare of upstream Niko1221/Strata baseline 296b59d2 (the clean fork reset on 2026-10-15) against upstream main reports 264 commits ahead / 0 behind. Compare: Niko1221/Strata@296b59d...main
This is a triage gate, not approval to merge 264 commits or to run a GPU build. Classify upstream changes touching CUDA prefill/MMQ, expert residency/dequant, KV/GDN/QSA/cache reuse, C1/C2 admission, MTP, sampling, vision and Linux stability. Exclude AMD-only/Metal/documentation-only work from the RTX 5080 hot-path shortlist except shared correctness dependencies.
Acceptance
- Record upstream head SHA, latest release and whether each relevant patch is merged/released/open/closed; dedupe against [P1][PI][RTX5080] Qualify C1/C2 concurrency for Pi on a single 16 GB RTX 5080 #2 [P2][PI][MEMORY] Validate C1↔C2 memory/admission policy on RTX 5080 #3 [P2][SPEC] Audit upstream probabilistic MTP and qualify Pi sampling benefit #10 [P1][PREFILL] Qualify 128K cold prefill and TTFT on RTX 5080 for Pi #11 [P1][CACHE] Qualify full-state 80K–120K Pi follow-up prefix reuse #13 [P1][PI][PREFILL] Profile short tool-result chunk dequant/launch hot path #18-[P2][RTX5080][EXPERT CACHE] A/B pinned arena and async tier, resolve conflicting 16GB prefill evidence #22.
- Summarize each candidate: exact commit/PR, current fork status, likely 5080 16GB Pi impact, evidence quality, correctness risks, and measurable test.
- Prioritize P0 [P1][RTX5080][Q6_K MMQ] Verify prefill speed and correctness on deployed UD-Q6_K_XL #19 correctness before throughput; run controlled short/32K/64K/~118K C1, C2 and Pi tool-loop A/B with fixed sampling, quant/pack, vision and KV flags.
- Selectively fast-forward or cherry-pick only after parity checks and an approved isolated benchmark window; preserve production NInfer service and clean upstream baseline until gate passes.
- If upstream itself fixes an existing issue, link and close/fold the obsolete fork issue instead of reimplementing.
No implementation or GPU benchmark is authorized by this issue alone.
Upstream ancestry realignment staging — 2026-10-20
- Preserved previous fork main (
1562c3347b9978c2875e32788fea066a5f59a707) asarchive/pre-ancestry-realignment-20261020. - Created
integration/upstream-ancestry-20261020from actual upstream commit ancestry at9ae93f7b1fe989d04a8e404d32cddaeeff10505f(upstream v0.1.46.1 era). - Reapplied PR fix(tools): enforce declared-tool boundaries across Strata API clients #16 declared-tool security patch as three branch-only commits, touching exactly
serve/frontend.py,serve/test_declared_tool_boundary.py, andserve/test_responses.py. GitHub comparison confirms branch = 3 ahead / 0 behind upstream anchor with precisely these three changed files. The patch hunks matched upstream source uniquely; runtime test results not yet available. - Existing fork
mainremains unchanged and GitHub confirms it has no common Git ancestor with this upstream-history branch. Therefore a normal internal PR/merge cannot repair main; use an explicitly reviewed, lease-guarded one-time ref transition after local CPU tests, open-PR/branch audit, production source isolation check and approval. Do not auto force-push, reset or deploy. - Next: test upstream-history branch's parser/API compatibility and namespace-alias regression suite in a disposable local checkout. Distinguish this ancestry maintenance effort from hot-path feature approval. Retain old history and all work branches.
Protection gate 2026-10-20
The attempted exact-SHA leased transition of main from 1562c3347b9978c2875e32788fea066a5f59a707 to upstream-aligned 2c9ad7e79933a7bab56ff6663dba95882cd3015f was rejected by GitHub protection: force pushes disabled and PR required. No update to main occurred. The old tip is archived, and candidate branch is prepared. A temporary, explicit owner-approved exception is required for the one-time unrelated-history transition, after which the original branch protections should be restored. Open PR #17 is based on old ancestry and needs separate recreation/rebase; do not delete its source branch. CPU rerun 48/48 passed; original run had an intermittent sampler lifecycle failure. Production stays pinned and untouched.
Final disposition — 2026-10-20 (completed)
The upstream ancestry migration was subsequently completed: fork main is 2c9ad7e79933a7bab56ff6663dba95882cd3015f, based on genuine upstream 9ae93f7b1fe989d04a8e404d32cddaeeff10505f. Verified upstream comparison reports exactly 3 ahead / 0 behind, preserving only the three PR #16 tool-boundary files. The previous fork main is archived at archive/pre-ancestry-realignment-20261020; the temporary Protect Main ruleset exception has been restored by the repository owner. CPU regression suite rerun: 48/48 passed (one earlier intermittent telemetry sampler test failure). No GPU benchmarks or deployment took place. The old PR #17 was auto-closed when history changed and its branch remains available; Issue #2 was closed/not planned with its harness preserved. This issue's prior 264-commit drift/sync/ancestry problem is now resolved at the repository level. Remaining performance experiments and fork-specific tests are independently tracked under #11, #18, #20-#22 and #24-#32. Closing this issue does not claim those GPU performance concerns were solved or production upgraded.
Activity
toddballinger commentedon Oct 17, 2026
Upstream PR hot-path shortlist (2026-10-17; inspected PR metadata, all OPEN at review)
P1 qualify first
P2 after C1/C2 and cache validation
Lower direct relevance: #2823 Q36 GGUF memory compaction saved ~18.5 GiB host RAM in a different Q36 engine, not 5080 GPU VRAM; #1942 remote vision useful only if remote encoder hardware exists.
No PR is recommended for blind merge. Check whether code has landed upstream before selectively transplanting; isolate actual Pi perf, 128K long context, production safe.
toddballinger commentedon Oct 18, 2026
2026-10-19 upstream hot-path follow-up — SM120 W4A4 and NVFP4 fork parity
Include the latest NVFP4 implementation evidence in the existing upstream selective-adoption gate:
Gate: current upstream/fork parity and correctness first (including P0 #19); then controlled RTX 5080 C1/C2 and short / 32K / 64K / genuinely ~118K occupied prompts with cold TTFT, 512/2048 decode, expert/VRAM/host traffic, Pi tool-loop correctness and vision tradeoffs. Fold results into existing #11/#22/#24/#25/#26 as appropriate. No new issue or unapproved merge/cherry-pick; preserve production baseline.
toddballinger commentedon Oct 19, 2026
Engineering audit checkpoint — 2026-10-19 (documentation only)
Verified: upstream
Niko1221/Stratabaseline296b59d2f73c8f04a805df29c2fa97655adace78→ upstreammain: GitHub compare reports 264 ahead / 0 behind. This repeats the count reported on Oct 17; it is not evidence of zero new commits, because the compare head may differ and must be pinned with SHA in the next audit. No blanket integration approved.Current fork inventory: 11 open issues (#2, #11, #18–#26), 5 closed issues (#3, #4, #5, #10, #13), 10 total PRs, 12 branches. PR #17 remains draft/open and mergeable; PR #16 merged; PR #15 closed unmerged. Closed #3, #10 and #13 still contain active Pi qualification language: explicitly reconcile state before reopening or folding into #2/#21; do not silently presume the work was completed.
Fresh upstream candidates for review, not verified improvements: Niko1221#3105 request-scoped prefill VRAM loans; Niko1221#3103 split MTP batch; Niko1221#3095 single-request pipeline windows in batch-mode; Niko1221#3109 SM89 benchmark reporting STRATA_DEQUANT_WIDE regression; Niko1221#3086 mapped expert-source prefetch. See https://github.com/Niko1221/Strata/pulls . All are upstream open PR candidates as seen October 19, not validated or merged changes.
Proposed dependency order: (1) #19 MMQ/tool-call correctness and #20 alias/graph failure, (2) #2/PR17 concurrency harness plus #25 task-quality parity, (3) #11 and #18 cold/short prefill, (4) #21 long-context state and #24/#26 vision + embeddings; #22 opt-in expert tuning only after baselines. Distinguish model/quant differences from engine effects. Record exact upstream head SHA, fork head SHA, quant pack/hash, flags, test provenance, measured useful completed tasks/minute, TTFT, RAM/VRAM and correctness before any cherry-pick.
Safety: this is a read-only research/documentation checkpoint. No GPU runs, CUDA builds, config changes, branch merges or production service changes authorized.
[-][P1][UPSTREAM][RTX5080] Triage 264 post-v0.1.44.1 commits and selectively qualify hot-path changes[/-][+][Other] [P1][UPSTREAM][RTX5080] Upstream ancestry and post-v0.1.44.1 sync[/+]