grenade

grenade commented on issue helexa/helexa#323

2026-09-07 11:32:15 +00:00

qwen4_exp gate 6: full-model logits parity against the upstream reference — the port is unverified against anything but our own reading

Source audit done, before the harness. One defect, one ambiguity settled.

The reference is now available — transformers from main (5.17.0.dev0) in a venv on beast reusing the system…

grenade pushed to main at helexa/helexa

2026-09-07 11:29:50 +00:00

d96e2fcfb7 Merge branch 'fix/323-qsa-scores-in-f32'

ea532c24dd fix(qwen4_exp): score QSA blocks in f32, as upstream does (#323)

Compare 2 commits »

grenade commented on issue helexa/helexa#311

2026-09-07 11:09:02 +00:00

qwen4_exp D: QSA sparse-attention indexer

The step-7 ambiguity is settled: our reading was right

Recorded here because this issue is where the question was raised, and it was raised as needing a reference trace. It did not — it…

grenade created branch fix/323-qsa-scores-in-f32 in helexa/helexa

2026-09-07 11:08:43 +00:00

grenade pushed to fix/323-qsa-scores-in-f32 at helexa/helexa

2026-09-07 11:08:43 +00:00

ea532c24dd fix(qwen4_exp): score QSA blocks in f32, as upstream does (#323)

grenade commented on issue helexa/helexa#316

2026-09-07 10:41:20 +00:00

validate-image.sh restores an evicted model with the wrong load spec, and the failure is silent

Two more guards a guarded window needs, both learned by breaking things

Five qwen4_exp cold-load windows on beast (2026-09-06/07) found two failure modes beyond the wrong-restore-spec this…

grenade opened issue helexa/helexa#323

2026-09-07 10:40:59 +00:00

qwen4_exp gate 6: full-model logits parity against the upstream reference — the port is unverified against anything but our own reading

grenade closed issue helexa/helexa#320

2026-09-07 10:38:31 +00:00

quant means two different things depending on the load path, so in-situ quantisation is unreachable on a single GPU

grenade commented on issue helexa/helexa#320

2026-09-07 10:38:27 +00:00

quant means two different things depending on the load path, so in-situ quantisation is unreachable on a single GPU

Fixed and confirmed live on beast (1ae3ba72).

Both halves of this issue are observable in the journal of a real cold load of the model that found it:

06:57:08  no GGUF in this repo;…

grenade commented on issue helexa/helexa#310

2026-09-07 07:08:50 +00:00

qwen4_exp C: host- or disk-resident PLE table with a device-side gather

First slice landed and validated against the real checkpoint

1ae3ba72 on main. The n-gram table is no longer made resident anywhere: the 128 shards are mapped where they already lie in…

grenade pushed to main at helexa/helexa

2026-09-07 06:22:48 +00:00

1ae3ba721e Merge branch 'feat/310-mmap-ngram-table'

7dfe18b1f5 feat(qwen4_exp): load the n-gram table first, so its failures are cheap

8c8d20c248 feat(qwen4_exp): map the n-gram table instead of making it resident (#310)

Compare 3 commits »

grenade pushed to feat/310-mmap-ngram-table at helexa/helexa

2026-09-07 05:46:21 +00:00

7dfe18b1f5 feat(qwen4_exp): load the n-gram table first, so its failures are cheap

grenade pushed to main at helexa/helexa

2026-09-07 05:44:19 +00:00

a79b9ae5ad Merge branch 'fix/quant-fallback-match-llamacpp'

15c664c5c3 fix(quant): use llama.cpp's fallback table, which spends bits rather than saving them

Compare 2 commits »

grenade created branch feat/310-mmap-ngram-table in helexa/helexa

2026-09-07 05:41:50 +00:00

grenade pushed to feat/310-mmap-ngram-table at helexa/helexa

2026-09-07 05:41:50 +00:00

8c8d20c248 feat(qwen4_exp): map the n-gram table instead of making it resident (#310)

grenade commented on issue helexa/helexa#307

2026-09-07 05:23:48 +00:00

Epic: serve Qwen3.8-Flash-Next (qwen4_exp) — a 180B sparse MoE whose architecture is built for beast's shape

Arithmetic correction: the experts are 73.0 GB, not 67.9

bf1e80ed claimed to meet the 640-wide reduction "the way llama.cpp meets the same wall". It did not. Upstream's tensor_type_fallback…

grenade commented on issue helexa/helexa#322

2026-09-07 05:23:33 +00:00

In-situ quantisation is redone on every load and discarded on unload — ~26 minutes per load for qwen4_exp

Two corrections and a boundary worth drawing before anyone designs this.

The GGUF claim above is wrong — withdraw it

I wrote that "our ISQ is not bit-identical to theirs and the substitution…

grenade pushed to fix/quant-fallback-match-llamacpp at helexa/helexa

2026-09-07 05:23:07 +00:00

15c664c5c3 fix(quant): use llama.cpp's fallback table, which spends bits rather than saving them

grenade created branch fix/quant-fallback-match-llamacpp in helexa/helexa

2026-09-07 05:23:05 +00:00

grenade commented on issue helexa/helexa#315

2026-09-07 05:08:18 +00:00

qwen4_exp H: quantisation, catalogue, fleet rollout and live validation

The residency paragraph above — "either pinned resident on beast or it is not served" — is presented as an operator decision, and #322 is why it currently is not one.

Measured on beast…