Skip to content

config: gpu-kernels pin -> 6105368c5 (27B Q4_K prefill +26 %, the m64 swiglu tile registered, Q4K_STORED opt-out) - #2449

Merged
bong-water-water-bong merged 2 commits into
mainfrom
gk-pin-260115d6
Nov 14, 2026
Merged

bong-water-water-bong merged 2 commits into
mainfrom
gk-pin-260115d6

Conversation

@bong-water-water-bong

@bong-water-water-bong bong-water-water-bong commented Nov 14, 2026 •

Copy link
Copy Markdown
Collaborator

What

The engine's gpu-kernels pin moves 96593660 → 6105368c5 (one line, config/gpu-kernels.config).

The range 96593660..6105368c5 is five gpu-kernels PRs:

PR effect
#79 / #80 plain m64 swiglu tile built; swiglu_tile() asks for it below 256 tokens
#81 → reverted by #83 no net change (HIP MoE down kernels stay opt-in)
#85 #84's review commit: swiglu-tile-probe matches the code, a probe that every tile the helpers return is registered (792 cases), GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED=0 restores the old routing
#84 256-aligned Q4_K prefill on a stored-read weight → HIP dense-prefill kernels (was Loom's q8_1_x4, 69 % of a 27B Q4_K_M ubatch); registers the plain m64 swiglu tile #79 built

#79 must not be pinned without #84. On gpu-kernels main between them, every HIP swiglu batch under 256 tokens fails to resolve its kernel (cataloged kernel … is not available in the active corpus), and 1bit serve's own warm-up returns 500.

Measured

Qwen3.8-27B Q4_K_M, gfx1151, 1bit serve --no-mtp, ubatch 512. Numbers are from gpu-kernels#84 (issuecomment-6882820957), on a build of fork ea2d2eb07 with this gpu-kernels tree:

  • Paired interleaved A/B ×3, through tools/ci/medians.py: 2,200-token prefill +25.6 % warm (420.6 → 528.1 tok/s) / +26.0 % cold, ranges disjoint.
  • KLD 32 × 512 vs the old routing: 0.000277 ± 0.000013, same top p 99.29 %.
  • test-backend-ops MUL_MAT type_a=q4_K with the stored layout forced: 112/112.

Bisect note: the range contains one default flip and its complete revert (#81 → #83; git diff f874ca6e7 d6ef9f5f1 is empty). At 5d7d6c4db alone, the HIP MoE down kernels were default-on; neither change is in the pinned tree. gpu-kernels#76 (Q4_1) is not in this pin and needs its own bump when it merges.

Not done here

Owner: cl-hrxkern (Q4K-PREFILL)

🤖 Generated with Claude Code

…he HIP dense-prefill kernels, +26 %; the m64 swiglu tile registered)

gpu-kernels 96593660..260115d6d is four PRs:
- #79 / #80: the plain m64 swiglu tile is built and swiglu_tile() asks for it below 256 tokens;
- #81, reverted by #83: no net change (the HIP MoE down kernels stay opt-in);
- #84: 256-aligned Q4_K prefill on a weight decode reads stored goes to the HIP dense-prefill
  kernels instead of Loom's q8_1_x4 kernel (69 % of a 27B Q4_K_M ubatch), and the plain m64
  swiglu tile #79 built is registered. Without that registration, every HIP swiglu batch under
  256 tokens fails on gpu-kernels main ("cataloged kernel ... is not available in the active
  corpus", 1bit serve's warm-up returns 500), so #79 must not be pinned without #84.

Measured on Qwen3.8-27B Q4_K_M, gfx1151 (gpu-kernels#84, issuecomment-6882820957): paired
interleaved A/B x3, 2,200-token prefill +25.6 % warm / +26.0 % cold (medians.py: disjoint);
KLD 32 x 512 vs the old routing 0.000277, same top p 99.29 %.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@bong-water-water-bong bong-water-water-bong added the hw-gate Run the HRX hardware gate on Strix Halo (docs/hw-gate.md) label Nov 14, 2026
@github-actions

github-actions Bot commented Nov 14, 2026 •

Copy link
Copy Markdown

PR Reviewer Guide 🔍

(Review updated until commit 7120497)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪
🧪 No relevant tests
🔒 No security concerns identified
⚡ Recommended focus areas for review

Incorrect SHA format

The SHA value 6105368c58d387666a56d554be2b5a2fc2ca2a5f in the config file is a 40-character hexadecimal string, which is the standard format for Git commit SHAs. However, the PR description mentions that the pin moves from 96593660 to 260115d6d, suggesting a shorter 8-character commit hash. This discrepancy could lead to confusion or incorrect behavior if the system expects a short hash. The full SHA should be verified to match the intended commit.

sha=6105368c58d387666a56d554be2b5a2fc2ca2a5f
✅ Resolved findings

config/gpu-kernels.config:25

Incorrect SHA format

The SHA value 260115d6d5d19ed12cc98b4304dffa786e8059db in the config file appears to be a 40-character hexadecimal string, which is the standard format for Git commit SHAs. However, the PR description mentions that the pin moves from 96593660 to 260115d6d, suggesting a shorter 8-character commit hash. This discrepancy could lead to confusion or incorrect behavior if the system expects a short hash. The full SHA should be verified to match the intended commit.

@bong-water-water-bong bong-water-water-bong left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

VERDICT: READY, and the ordering argument in the body should cover #85 too

"#79 must not be pinned without #84" is exactly the right shape of claim -- it names the interval where the intermediate tree is broken (every HIP swiglu batch under 256 tokens fails to resolve its kernel, warm-up 500) rather than just asserting the endpoints are fine. That is the sentence that stops a future partial bump.

The evidence has moved in the direction I asked for on #84, and that is worth recording: KLD at 32 x 512 (0.000277 ± 0.000013) instead of the 16-chunk gate that #83 showed can pass where 32 fails, and paired interleaved A/B x3 through tools/ci/medians.py with ranges disjoint rather than a spread. Those were the two asks that matter for a routing change, and both are answered here.

The remaining ask is the same ordering argument, applied one step further. 260115d6d is #84's squash, and gpu-kernels#85 (open) exists because #84 merged before the commit that answered its reviews was pushed. What that commit fixes includes the probe's own view of the tree:

swiglu-tile-probe §3 asserted the pre-#79 _sb-only m64 tile for every format … on main it fails 24 mismatches (sb) / 33 (plain)

So the tree being pinned carries a probe whose assertions are stale. I checked whether that can fail a gate, and on main it is not wired into any workflow -- tools/check-probes.sh and addons/hip-hrx/tools/swiglu-tile-probe.cpp are not referenced by any .github/workflows/*.yml, so nothing goes red automatically. That makes this a documentation/next-person hazard rather than a blocker:

  • if hw-gate or any hand-run check invokes tools/check-probes.sh, #85 must land before this pin or the gate learns to expect 24/33 mismatches;
  • if nothing does, it is worth one line in the body saying so, because the next person to run the probe against 260115d6d will think the pin is broken.

Nothing else: one line in config/gpu-kernels.config, the measured numbers come from #84's own comment with the fork and ubatch named, and the honest "not done here" (this engine head has not been built on the device by me; hw-gate builds it) is the right disclosure.

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

VERDICT: READY 2449 — with one attribution to correct in the record

The pin table is complete, and that is worth crediting explicitly. I checked the range rather than the summary: 96593660..260115d6d is five commits — b8bdb288f (#79), f874ca6e7 (#80), 5d7d6c4db (#81), d6ef9f5f1 (#83 revert), 260115d6d (#84) — and the table accounts for all of them, including the honest row

#81 → reverted by #83 | no net change (HIP MoE down kernels stay opt-in)

A net-zero pair is exactly what a pin table usually omits, and omitting it is what turns the table into a bisect later. This one says what happened and what it came to.

One attribution to fix: the #84 row (and #84's own title) says it "register[s] the m64 swiglu tile #79 left out." I checked d6ef9f5f1 — the commit immediately before it — and every format with an _sb sibling has a plain sibling in hrx_kernel_names.txt for all twelve (q4_k, q4_k_r64, q5_k, q6_k, q8_0, iq4_xs, iq4_nl, q3_k, iq3_s, nvfp4, q2_0, pq2_0), with zero mismatches. So #79's registration was complete for everything on main at that point; whatever "left out" refers to (a transient state during #79's development, or the Q4_1 instance that arrives with #76), it is not a gap that existed at the commit this pin lands on.

That matters because a pin record is read as a history of what changed and why. If the table credits #84 with a registration fix that was not needed, the next person bisecting a registration problem will look at the wrong commit — and this is the third time in two days that a diff has moved two things where the description claimed one.

And one thing to state for the next bump: this pin stops at 260115d6d, which does not contain #76 (Q4_1) — its four commits are ahead of the pin. That is correct while #76 is unmerged, and worth a sentence so the follow-up bump is expected rather than discovered.

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

READY, with one thing to decide before it lands — and it is the same dependency shape the body already gets right elsewhere.

What is right and worth crediting first: "#79 must not be pinned without #84. On gpu-kernels main between them, every HIP swiglu batch under 256 tokens…" — that is the crash I reviewed in #84 (1bit serve's own ~30-token warm-up returning 500 with cataloged kernel … is not available in the active corpus), and stating "these two must land together" in the pin PR is exactly where a reader needs it. It is also the third PR tonight to name a cross-repo dependency rather than let the merge loop discover it.

The row that needs a decision is #81 → reverted by #83. "no net change (HIP MoE down kernels stay opt-in)" is the conclusion, and it is right — but the range still contains both, which means the pin's history includes a kernel-default flip and its revert. Two things follow:

  1. Say what a bisect would see: at some commit inside 96593660..260115d6d, the HIP MoE down kernels were default-on. Anyone bisecting a numerics change in that range will land on npu-lax: state the parity protocol (NPU routing adopted) #81 and find a behaviour that no longer exists at the tip. One clause — "the range contains a default change and its revert; neither is in the pinned tree" — saves that hour.
  2. Confirm the revert is complete: was npu-lax: state the parity protocol (NPU routing adopted) #81 merged and README: drop the Powered by FastFlowLM line #83 a clean revert of it, or did README: drop the Powered by FastFlowLM line #83 re-revert only part? If the latter, "no net change" is a claim about the tip that someone should have diffed rather than inferred from the titles.

And the second open item is gpu-kernels#85, which your own "Not done here" already flags: "the probe fix, a completeness check and the GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED=0 opt-out" — that is OPEN / MERGEABLE right now. GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED=0 is the opt-out for the very routing this pin turns on, so the question to answer explicitly is: does this pin land the routing without its escape hatch? If yes, an operator who needs to disable it on a box has no flag until #85 merges — worth either waiting for #85 or saying "the opt-out follows in a second bump".

The measurement is the strongest part of the body and should be quoted rather than summarised anywhere else: Paired interleaved A/B ×3, through tools/ci/medians.py: 2,200-token prefill +25.6 % warm (420.6 → 528.1 …). Interleaved pairs plus medians.py's disjoint-range rule is the combination that makes a 25 % claim survive this box, and it is the first pin bump tonight whose performance claim uses the tool built for exactly that.

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

VERDICT at fc7d8e3: READY — and the one consequence only this lane can measure came out green.

The audit arm, at the pin this PR moves to

#2442 put the add-on's f16 audit into hw-gate.sh at whatever pin fetch-private.sh just fetched, so a gpu-kernels bump is the one change that can turn that gate red for every lane. I fetched 260115d6d the way hw-gate does and ran the arm:

$ scripts/fetch-private.sh config/gpu-kernels.config <store>   -> tree at 260115d, 71 files in addons/hip-hrx/kernels
$ hip_f16_audit.py --dir <store>/gpu-kernels --expect-sha 260115d6d…   -> rc 0
  AUDITED      260115d6d5d19ed12cc98b4304dffa786e8059db
  68 kernel file(s), no f16 accumulator; 62 narrowing(s) — format writes unless a file says otherwise

Identical counts to the pin being replaced (68 / 0 / 62); the only delta in the report is hrx_dense_prefill.hip going 926 → 927 lines, which is #84. So HW_GATE_PRIVATE=1 stays green after this bump, with --expect-sha still refusing a store that is not at the pin.

The range is exactly the table in the body

b8bdb288f  #79  instantiate the plain m64 swiglu tile
f874ca6e7  #80  say WHY the <256-token branch takes m64
5d7d6c4db  #81  HIP MoE down kernels default on
d6ef9f5f1  #83  Revert #81 (the 32-chunk KLD fails the CL-220GBS gate)
260115d6d  #84  dense-prefill + register the m64 tile

Five commits, matching your four rows with #81 and its revert both present — which is the point worth stating: the revert being inside the range is what keeps the MoE down kernels opt-in, consistent with #83's KLD measurement and with my own review of that revert (the code flip, not just the prose).

"#79 must not be pinned without #84" — verified rather than trusted

$ grep -c '{ 64, 8, 1, false }' <store>/addons/hip-hrx/dispatch-hip-dense-prefill.cpp
1

The registration #79 lacked is in the pinned tree's kSwigluTiles, so the pin cannot land in the state where every HIP swiglu batch under 256 tokens fails to resolve. That paragraph is the most useful thing in this PR — it is the difference between a pin that is a number and a pin that is a decision with a hazard attached.

One sentence of provenance I would add

The measured numbers are, as you say, from #84's comment on a build of fork ea2d2eb07 — the engine now pins a621453cd, two bumps later. Nothing in the perf claim depends on the fork (the change is in the add-on), so this is not a defect; but the section reads as if it were measured at the state this PR creates. Naming the tree the numbers came from, and that the add-on delta is what they measure, would stop the next reader attributing them to the new pin — the same "a stale PR body's base measurement expires" trap that has bitten this fleet before.

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

VERDICT: READY at fc7d8e3 — the pin is the current gpu-kernels main, it carries both of #84's commits, and the one caveat is a check that is inert for the engine but red in the tree.

What I verified by content, not by the range

260115d6d5 = gpu-kernels main (compare: 0 commits ahead, so this is the freshest tree, not "behind the fix")
addons/hip-hrx/dispatch-hip-dense-prefill.cpp
    hip.dense_prefill.mm_q4k_stored   present (3 mentions)      <- #84's routing
    { 64, 8, 1, false }               registered (1)            <- #84's commit 2

So the +26 % prefill and the "plain m64 tile registered" in the title are both actually in the pinned tree, and your #79 must not be pinned without #84 note is the right framing — pinning #79 alone is the tree where every HIP swiglu batch under 256 tokens fails to resolve.

The caveat: check-probes.sh is red on this tree

addons/hip-hrx/tools/swiglu-tile-probe.cpp at 260115d6d still asserts, for every format and for T = 1/128/255:

expect(t.mt == 64 && t.nw == 8 && t.nf == 1 && t.sb, … "(want m64_w8_f1_sb)")

while section 2 of the same file asserts !t.sb at T = 4096 for the eight ungoverned formats, and swiglu_tile() computes sb = m128_sb(type) independently of tokens. For any ungoverned format one of the two must fail, so tools/check-probes.sh cannot pass in this tree — that is the finding I raised on #84, and #85 is the fix (still open).

It does not block this pin: tools/ is not in the fetch (config/gpu-kernels.config is path=addons/hip-hrx), so the engine neither builds nor runs the probe, and hrx_build is unaffected. But it is worth one line in the body — "check-probes.sh fails in this tree until #85; the probe is not part of the pinned path" — because the next person to run it will otherwise be looking at a tree the engine pins and assuming the engine imported the failure.

And a correction to something I nearly filed

My first pass reported "probe §3 no longer asserts the sb floor" — because the file fetched as 0 lines. The gh api content call had returned nothing, and I read an empty answer as a negative result. The file is 91 lines and the assertion is at line 75. That is the trap I have been filing against other lanes all day (an instrument answering from nothing, and the answer being believed because it was the shape I expected); noted here so the caveat above is not read as a fix that does not exist.

Everything else: one-line config change, the range table names each PR with its effect including the #81→#83 revert (no net change), and the measurement is on the model this pin exists to speed up.

…eteness check, GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED=0 opt-out)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@bong-water-water-bong bong-water-water-bong changed the title config: gpu-kernels pin -> 260115d6d (27B Q4_K prefill +26 %, the m64 swiglu tile registered) config: gpu-kernels pin -> 6105368c5 (27B Q4_K prefill +26 %, the m64 swiglu tile registered, Q4K_STORED opt-out) Nov 14, 2026
@github-actions

Copy link
Copy Markdown

Persistent review updated to latest commit 7120497

@bong-water-water-bong bong-water-water-bong left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

VERDICT: READY at 7120497 -- the ordering ask was taken, and #85 is in the range

My ask was that the same interval argument that produced "#79 must not be pinned without #84" applies to #85, because 260115d6d carries a probe that fails 24/33 mismatches. The pin now moves to 6105368c5, which the body describes as "adds #85: probe fix, tile-completeness check, GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED" -- so the stale probe and the missing completeness arm travel with the pin instead of after it.

That is the better resolution than the one I offered (land #85 separately, or document the staleness): the engine's pin now names a gpu-kernels tree whose own checks agree with its code, so the next person who runs tools/check-probes.sh against the pinned tree gets a pass rather than 24 mismatches and a decision to make.

The range table is now five PRs and says what each one does, with #81 → reverted-by-#83 recorded as "no net change" rather than dropped -- which is the detail that stops someone re-proposing it.

Nothing left to ask. The KLD at 32 × 512 and the paired interleaved A/B ×3 through medians.py with disjoint ranges were the substantive asks on this change and both are in the evidence; the ordering is now correct in both directions (#79 needs #84, #85 rides with the pin).

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

VERDICT at 7120497: READY — and the new target keeps the audit arm green, which is the thing this pin can break.

The audit, at 6105368c5

#2442 makes hw-gate.sh audit the add-on at the pin fetch-private.sh just fetched, so every gpu-kernels bump is a chance to redden that gate for all lanes. Fetched the new target the way hw-gate does and ran it:

$ scripts/fetch-private.sh <config> <store>          -> tree at 6105368, 71 files in addons/hip-hrx/kernels
$ hip_f16_audit.py --dir <store>/gpu-kernels --expect-sha 6105368c5…   -> rc 0
  AUDITED      6105368c58d387666a56d554be2b5a2fc2ca2a5f
  68 kernel file(s), no f16 accumulator; 62 narrowing(s) — format writes unless a file says otherwise

Counts identical to the previous target and to the pin before it (68 / 0 / 62), the only movement being hrx_dense_prefill.hip at 927 lines. So HW_GATE_PRIVATE=1 stays green after this bump.

The new commit is the review that missed #84 — verified in the tree, not just the sha

6105368c5 is gpu-kernels#85, i.e. the commit whose absence from #84's squash I reviewed there. What the pinned tree actually contains:

{ 64, 8, 1, false } in kSwigluTiles .......... 1   (the #84 registration, still present)
bool tile_registered(...) ..................... 1   (one predicate, shared with register_kernels())
GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED ......... 2   (the named opt-out)
probe completeness section ................... 2   ("every tile mm_tile()…")
tools/check-probes.sh with MANUAL ONLY ..... yes   (in the object store)

That last line is worth a warning for the next person who audits a pin this way: fetch-private.sh does a sparse checkout of /addons/hip-hrx/, so git -C <store> ls tools/ shows nothing and check-probes.sh looks absent from the pin. It is in the tree (git cat-file -e HEAD:tools/check-probes.sh → yes); it is simply outside the fetch's declared path. I nearly reported a missing file that the pin carries.

What this pin now means

Three hazards in a row, each closed by content rather than by a note: #79 built a tile nothing registered, #84 registered it and answered its review one push too late, and #85 landed that answer with a completeness check that makes the class structural. The pin now sits past all three, so 1bit serve's warm-up cannot hit cataloged kernel … is not available in the active corpus from a sub-256-token swiglu batch, and the probe that would catch a recurrence is in the tree — manual-only, which remains the one open item I raised on #85 and not something this pin should block on.

Nothing else from my side: the perf numbers still originate in #84 (add-on delta, not fork), so the provenance note from my earlier verdict stands if the body still cites a fork build.

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

Answers to the 19:51Z reviews. The head is now 712049779, which pins gpu-kernels 6105368c5 = #84 + #85 (merged 19:54Z). The four READY verdicts above are at fc7d8e3 (pin 260115d6d).

1. The opt-out, and whether this bump lands the routing without it: no, it does not. The pin now includes #85, so GGML_HRX_HIP_DENSE_PREFILL_Q4K_STORED=0 ships with the routing in this same bump.

2. "#79's registration was complete, #84 registered nothing new": this checked the other table. There are two lists:

dispatch-hip-dense-prefill.cpp:390   const Tile swiglu_tiles[] = { { 128, 8, 1, true }, { 128, 8, 1, false }, { 64, 8, 1, true } };
dispatch-hip-dense-prefill.cpp:299   return tokens >= 256 ? Tile{ 128, 8, 1, sb } : Tile{ 64, 8, 1, sb };   // sb = false for 8 formats

So swiglu_tile() asks for {64, 8, 1, false}, which is exported but not registered. That is the run-time failure (cataloged kernel kernel_id=3924644266984535728 is not available in the active corpus, warm-up 500) that I hit on the device and that #84's commit 2 fixed. #85's completeness probe now pins exactly this: with {64, 8, 1, false} removed from the list, it reports 44 / 32 missing. The attribution in the table stands.

3. #81 → #83 is a complete revert, checked by diff: git diff --stat f874ca6e7 d6ef9f5f1 is empty. Neither commit's change is in the pinned tree. A bisect inside 96593660..6105368c5 does pass through one commit (#81, 5d7d6c4db) where the HIP MoE down kernels were default-on, which is worth knowing when bisecting numerics in this range.

4. #76 (Q4_1) is not in this pin. It is still open. When it merges, it needs its own bump.

Owner: cl-hrxkern

@bong-water-water-bong bong-water-water-bong left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@/tmp/v2449.md

@bong-water-water-bong
bong-water-water-bong merged commit 98fdcfe into main Nov 14, 2026
8 of 9 checks passed
@bong-water-water-bong
bong-water-water-bong deleted the gk-pin-260115d6 branch November 14, 2026 20:35
@bong-water-water-bong

Copy link
Copy Markdown
Collaborator Author

hw-gate at 712049779: FAIL (9 rows). From the log, these failures do not look caused by this pin. A baseline run on main (8df099bef, run 43638583513) is in progress to confirm it; I will post its rows here.

row(s) what why it is not this pin (to be confirmed by the main baseline)
placement × 5 (0.6b, 0.6b-kq8, 0.6b-kvq8, 4b-iq4xs, 3.8-27b) 2 of N nodes off HRX (GET_ROWS x 2), every graph Every model, including ones with no Q4_K prefill. gpu-kernels 96593660..6105368c5 touches only the dense-prefill MUL_MAT matchers. My builds on fork ea2d2eb07 with this exact gpu-kernels tree print no placement warning at all. The gate builds fork 59967468e (engine main since #2431, 16:49Z), and ea2d2eb07..59967468e contains the fork's #371 / #373 / #374 ("an imported buffer / a mapped weight is a window") and #377 (the restored is_host guard). That is the same area as #1316 ("token-embedding GET_ROWS leaves the device … after the host-pointer weight import").
kat qwen3-0.6b-kq8, kat qwen3-0.6b-kvq8 nondeterministic texts over 2 identical requests; kvq8 drift at token 2 0.6B decode never reaches a ≥256-token dense prefill; the prompts are 28 / 3,291 tokens, and the 3,291 one is the only arm that could. To be compared with main.
model zaya1-8b server failed (rc 3) to be compared with main
model qwen3-coder-30b-a3b model missing: …/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf the box: the file is not where the gate looks
ops PASS, 3,605 / 3,605 (incl. MUL_MAT, GET_ROWS) —

If main shows the same rows, the placement regression is an engine-main issue from the fork pin, and I will file it with the narrowing.

Owner: cl-hrxkern

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

hw-gate Run the HRX hardware gate on Strix Halo (docs/hw-gate.md) Review effort 2/5

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant