Bump @types/node from 22.16.5 to 24.10.0#165
Closed
dependabot[bot] wants to merge 1 commit into
Closed
Conversation
Bumps [@types/node](https://github.com/DefinitelyTyped/DefinitelyTyped/tree/HEAD/types/node) from 22.16.5 to 24.10.0. - [Release notes](https://github.com/DefinitelyTyped/DefinitelyTyped/releases) - [Commits](https://github.com/DefinitelyTyped/DefinitelyTyped/commits/HEAD/types/node) --- updated-dependencies: - dependency-name: "@types/node" dependency-version: 24.10.0 dependency-type: direct:development update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com>
dependabot
Bot
added
dependencies
Contributor
Author
|
OK, I won't notify you again about this release, but will get in touch when a new version is available. If you'd rather skip all updates until the next major or minor version, let me know by commenting If you change your mind, just re-open this PR and I'll resolve any conflicts on it. |
dependabot
Bot
deleted the
dependabot/npm_and_yarn/types/node-24.10.0
branch
joelteply
added a commit
that referenced
this pull request
Apr 8, 2026
…tillation-first pivot
Three new subsections in §4.1, capturing tonight's negative results across
the dense-model forge branch and proposing the structural reframe that the
empirical work converged on.
§4.1.4.2 — Per-tier Pareto comparison (negative result):
The aggressive-quantization Pareto test from §4.1.4.1 is now run end-to-end.
v2-7B forge and unmodified Qwen2.5-Coder-7B base, both quantized to Q5_K_S /
Q3_K_M / Q2_K via the same llama.cpp toolchain, both evaluated via the
same patched vLLM-GGUF backend that anchored against Qwen's published
61.6/53.0 to within +0.6/+0.7 (deterministic across 5+ runs):
Tier v2-7B base Δ Winner
Q5_K_S 5.0G 55.5 63.4 -7.9 base by +14% on quality/vram
Q3_K_M 3.6G 54.3 59.8 -5.5 base by +10% on quality/vram
Q2_K 2.9G 42.7 43.3 -0.6 tie within run noise
Base+quant Pareto-dominates the v2-7B forge at every tier we tested. The
closest the forge gets to parity is Q2_K (within run noise). By the
§4.1.4.1 product-relevance criterion, the v2 forge methodology as
currently constituted does not produce a useful product on the
Qwen2.5-Coder-7B family.
Three independent failure modes ruled out as fixes for the residual gap,
all on the same v2-7B base, all with disciplined cause-of-the-gap
comparison:
1. More cycles + more training: WORSE (54.9 → 46.3)
2. Held-out-aware code calibration: NO IMPROVEMENT (54.9 → 53.7)
3. Aggressive quantization: NO (base+quant wins at every tier, this section)
Each was the leading hypothesis when tested. Each was falsified. Three
negative results in succession on three independent fix candidates is
strong empirical evidence that the activation-magnitude head-pruning +
LoRA-recovery approach does not have a Pareto-improving sweet spot for
dense base models that already have good quantization options,
regardless of which knob in the strategy space is tuned.
§4.1.4.3 — Product positioning implications:
Two non-overlapping product positions where the forge has a defensible
value proposition, and dense Qwen2.5-Coder-7B-class models with good
Q3/Q2 quantizations are not one of them:
1. Distillation-first compaction (any base model, dense/MoE/hybrid)
— see §4.1.5 below
2. Pre-removable expert pruning + structural compaction (MoE/hybrid/
oversized targets that base+quant cannot reach at all because the
base does not fit on the target hardware even at Q2_K)
The Qwen3.5-35B-A3B (target A) and Qwen3.5-397B-A17B (target B grid
moonshot) work falls in product position #2. Their value is making
models reachable that the alternative compaction methods cannot reach,
not matching a base+quant alternative at the same tier.
The dense-model forge work is suspended until distillation-first lands.
§4.1.5 — Distillation-first compaction (the next-iteration methodology
proposal):
The empirical pattern across §4.1.3.1, §4.1.3.2, §4.1.4.1, and §4.1.4.2
is consistent enough to motivate a structural pivot in the forge's
primary compaction mechanism, and the substrate work to support that
pivot has already landed (compensation_lora.py + test_compensation_lora.py
+ COMPENSATION-LORA-DESIGN.md, all committed in a previous PR commit).
The pivot inverts the v2 dependency: instead of "structured prune +
LoRA recovery against fine-tuning loss", the new mechanism is "any
transformation + distill against the unmodified teacher's hidden states".
Pruning, quantization, modality fusion, context extension all become
slot-in transformations that the (transform, distill, eval) loop
recovers from independently. The methodology becomes a search procedure
over the transformation space, with distillation as the convergence step
and the per-tier metric as the stopping criterion.
The smoke test on distilgpt2 passed all five stability checks
(tokenizer alignment, hidden-state magnitudes within 2× across layers,
loss decreased monotonically -39.35% relative, per-layer losses balanced
within 3.14× of median, no NaN/inf). The math is sound at small scale;
production scale-up to 7B is unblocked by RUNNING the script, not by
writing more code.
§4.1.5 results paragraph queued: take the v2-7B artifact from row 5,
apply compensation LoRA with Qwen2.5-Coder-7B base as teacher and a
held-out-aware calibration mixture, measure HumanEval through the
same calibrated pipeline. Success criterion: HumanEval pass@1 ≥ 58.0
(a 3-point improvement, just outside the calibration tolerance band).
If at-or-above: distillation-first is empirically validated, dense-model
forge branch unfreezes, §4.1.4 row 5 gets a successor row. If below:
follow the failure escalation path documented in COMPENSATION-LORA-DESIGN.md
to the cross-layer skip path.
Independence from the moonshot work is explicit: distillation-first is
the dense-model branch's path forward, MoE/hybrid expert pruning is a
separate substrate path, the two branches advance in parallel.
Plus durability adds:
- docs/hf-deprecation-notices/qwen2.5-coder-14b-compacted-DEPRECATION.md
(the user-facing model card replacement for the v1 deprecated artifact)
- docs/papers/NEUROPLASTIC-SUBSTRATE.md
(the architectural-thesis paper from earlier in the work session)
References:
- sentinel-ai PR #161 (the substrate work that produced these results)
- sentinel-ai#160 / #161 / #163 / #164 / #165 (the issues these subsections
document the resolution paths for)
- Qwen2.5-Coder Technical Report Table 5 (the anchor for the per-tier
measurements)
joelteply
added a commit
that referenced
this pull request
Apr 8, 2026
* Paper stub: Validated Structured Pruning for Consumer Hardware Companion to Experiential Plasticity. Documents the layered test harness for tensor surgery validation. Two real bugs in production pruning code were caught during harness construction: 1. LoRA-on-pruned-hooks corrupts model on hook removal 2. Config drift after defrag breaks save/load roundtrip Title, abstract, and section outline. Full paper to follow as Layers 4-6 of the harness are built. * Paper: bugs table and Layer 4 update The harness now caught 5 real bugs during construction: 1. LoRA-on-pruned-hooks corruption (FIXED) 2. Config drift after defrag (FIXED) 3. Hybrid attention models (TRACKED) 4. L2 norm importance is unreliable (RESEARCH FINDING) 5. Pruning without retraining is destructive (RESHAPES the experiential plasticity narrative — recovery comes from fine-tuning, not pruning) * Papers: write down 5 findings into both experiential and validated-pruning Experiential paper: - New section 9.5: Validation and a Reframing of the Plasticity Story - Documents LoRA-on-pruned-hooks bug that produced phantom +88% improvements - Documents L2-norm importance metric finding (anti-correlated with importance) - Reframes the central claim: recovery comes from fine-tuning, not smart pruning - Calls for no-prune equal-budget fine-tune baseline as required ablation - Adds reference to companion paper Validated pruning paper: - New section: Findings In Detail (5 findings, each with empirical evidence) - Each finding has the failure mode, the empirical signature, and the fix - Frames validation harnesses as required artifacts for pruning papers * Papers + kash-feedback: gpt2 OUTCOME D, four-metric finding, Layer 6 EXPERIENTIAL-PLASTICITY: - New §9.5 subsection: gpt2-medium re-run on corrected pipeline - Reframes §4 transfer function as "early-cycle behavior controller will encounter" - Documents that the controller's quality-aware stopping criterion makes the cycle 9 anomaly structurally unreachable in production - Adds OUTCOME D framing: closed-loop controller enforces the transfer function - Adds four-metric comparison subsection with the activation > saliency >> L2 > gradient ranking - Calls out PLASTICITY-COMPACTION's gradient-magnitude trick as a publication-blocking question VALIDATED-TENSOR-SURGERY: - New section: Layer 6, the structural fix that closes the bug class - Documents the literal 62→7→501 historical bug pattern catch - New section: The four-metric comparison empirical result - Hypothesizes why activation alone beats saliency at small calibration - Recommends activation-magnitude as the default forge pipeline metric KASH-FEEDBACK.md: - Appended results from both experiments - OUTCOME D framing for gpt2 result - Three observations on the four-metric finding - Three questions for Kash's read Refs continuum #841 (gpt2 re-run, partial OUTCOME D) Refs continuum #842 (Layer 6 invariant, shipped) Refs continuum #844 (four-metric comparison, surprise result) Refs sentinel-ai #155 (importance metric finding, now contextualized) * PLASTICITY-COMPACTION: §4.1.4.2 per-tier negative result + §4.1.5 distillation-first pivot Three new subsections in §4.1, capturing tonight's negative results across the dense-model forge branch and proposing the structural reframe that the empirical work converged on. §4.1.4.2 — Per-tier Pareto comparison (negative result): The aggressive-quantization Pareto test from §4.1.4.1 is now run end-to-end. v2-7B forge and unmodified Qwen2.5-Coder-7B base, both quantized to Q5_K_S / Q3_K_M / Q2_K via the same llama.cpp toolchain, both evaluated via the same patched vLLM-GGUF backend that anchored against Qwen's published 61.6/53.0 to within +0.6/+0.7 (deterministic across 5+ runs): Tier v2-7B base Δ Winner Q5_K_S 5.0G 55.5 63.4 -7.9 base by +14% on quality/vram Q3_K_M 3.6G 54.3 59.8 -5.5 base by +10% on quality/vram Q2_K 2.9G 42.7 43.3 -0.6 tie within run noise Base+quant Pareto-dominates the v2-7B forge at every tier we tested. The closest the forge gets to parity is Q2_K (within run noise). By the §4.1.4.1 product-relevance criterion, the v2 forge methodology as currently constituted does not produce a useful product on the Qwen2.5-Coder-7B family. Three independent failure modes ruled out as fixes for the residual gap, all on the same v2-7B base, all with disciplined cause-of-the-gap comparison: 1. More cycles + more training: WORSE (54.9 → 46.3) 2. Held-out-aware code calibration: NO IMPROVEMENT (54.9 → 53.7) 3. Aggressive quantization: NO (base+quant wins at every tier, this section) Each was the leading hypothesis when tested. Each was falsified. Three negative results in succession on three independent fix candidates is strong empirical evidence that the activation-magnitude head-pruning + LoRA-recovery approach does not have a Pareto-improving sweet spot for dense base models that already have good quantization options, regardless of which knob in the strategy space is tuned. §4.1.4.3 — Product positioning implications: Two non-overlapping product positions where the forge has a defensible value proposition, and dense Qwen2.5-Coder-7B-class models with good Q3/Q2 quantizations are not one of them: 1. Distillation-first compaction (any base model, dense/MoE/hybrid) — see §4.1.5 below 2. Pre-removable expert pruning + structural compaction (MoE/hybrid/ oversized targets that base+quant cannot reach at all because the base does not fit on the target hardware even at Q2_K) The Qwen3.5-35B-A3B (target A) and Qwen3.5-397B-A17B (target B grid moonshot) work falls in product position #2. Their value is making models reachable that the alternative compaction methods cannot reach, not matching a base+quant alternative at the same tier. The dense-model forge work is suspended until distillation-first lands. §4.1.5 — Distillation-first compaction (the next-iteration methodology proposal): The empirical pattern across §4.1.3.1, §4.1.3.2, §4.1.4.1, and §4.1.4.2 is consistent enough to motivate a structural pivot in the forge's primary compaction mechanism, and the substrate work to support that pivot has already landed (compensation_lora.py + test_compensation_lora.py + COMPENSATION-LORA-DESIGN.md, all committed in a previous PR commit). The pivot inverts the v2 dependency: instead of "structured prune + LoRA recovery against fine-tuning loss", the new mechanism is "any transformation + distill against the unmodified teacher's hidden states". Pruning, quantization, modality fusion, context extension all become slot-in transformations that the (transform, distill, eval) loop recovers from independently. The methodology becomes a search procedure over the transformation space, with distillation as the convergence step and the per-tier metric as the stopping criterion. The smoke test on distilgpt2 passed all five stability checks (tokenizer alignment, hidden-state magnitudes within 2× across layers, loss decreased monotonically -39.35% relative, per-layer losses balanced within 3.14× of median, no NaN/inf). The math is sound at small scale; production scale-up to 7B is unblocked by RUNNING the script, not by writing more code. §4.1.5 results paragraph queued: take the v2-7B artifact from row 5, apply compensation LoRA with Qwen2.5-Coder-7B base as teacher and a held-out-aware calibration mixture, measure HumanEval through the same calibrated pipeline. Success criterion: HumanEval pass@1 ≥ 58.0 (a 3-point improvement, just outside the calibration tolerance band). If at-or-above: distillation-first is empirically validated, dense-model forge branch unfreezes, §4.1.4 row 5 gets a successor row. If below: follow the failure escalation path documented in COMPENSATION-LORA-DESIGN.md to the cross-layer skip path. Independence from the moonshot work is explicit: distillation-first is the dense-model branch's path forward, MoE/hybrid expert pruning is a separate substrate path, the two branches advance in parallel. Plus durability adds: - docs/hf-deprecation-notices/qwen2.5-coder-14b-compacted-DEPRECATION.md (the user-facing model card replacement for the v1 deprecated artifact) - docs/papers/NEUROPLASTIC-SUBSTRATE.md (the architectural-thesis paper from earlier in the work session) References: - sentinel-ai PR #161 (the substrate work that produced these results) - sentinel-ai#160 / #161 / #163 / #164 / #165 (the issues these subsections document the resolution paths for) - Qwen2.5-Coder Technical Report Table 5 (the anchor for the per-tier measurements) * PLASTICITY-COMPACTION §4.1.3.4 + forge template architecture rule §4.1.3.4: The importance-metric calibration lesson generalizes across structural unit (heads → experts). Empirical anchor: continuum-ai/qwen3-coder-30b-a3b-compacted-19b-256k v1 (alloy hash aa61c4bdf463847c). Hardware-measured 88.4 HumanEval / 86.0 HumanEval+ against unmodified Qwen3-Coder-30B-A3B-Instruct base anchor at 92.1 / 89.0, both on RTX 5090 + llama.cpp Q5_K_M in the same eval pipeline. The artifact carries the router-gate-L2-norm baseline (78.7 HumanEval) in priorMetricBaselines[] as the negative-baseline empirical control that makes the §4.1.3.4 claim falsifiable from the published artifact alone. The structural lesson: the metric-calibration pattern from §4.1.3.1 (dense head pruning) recurs at the MoE expert level. Router-gate-L2-norm is the architectural-only equivalent of activation-magnitude-only head importance, and replacing it with calibration-aware activation counts on a held-out code corpus closes +9.7 HumanEval points on the same prune budget, no fine-tuning, no compensation. ANY importance metric for ANY prunable unit (heads, experts, layers, future structural units) must be derived from task-conditioned activation profiling on a held- out corpus that reflects the artifact's intended workload. The two data points form a methodology curve, not a single anomaly. The compensation v2 step (KL distillation on top of the calibration- aware student to push from 88.4 → projected 90+) is currently blocked on transformers' caching_allocator_warmup pre-allocating an fp16 buffer equal to full model size before bnb 4-bit quantization takes effect, exceeding total VRAM on a single 32 GB GPU even with both teacher and student nominally 4-bit. The architecturally correct fix is offline teacher-logit precomputation (phase 1: load teacher alone, dump logits; phase 2: unload; phase 3: load student alone, train against on-disk logits). This is the next sentinel-ai sprint and is documented in §4.1.3.4's "next experimental wave" paragraph. CLAUDE.md: Forge Template Architecture rule. The qwen3-coder-30b-a3b-compacted-19b-256k v1 publish required ~6 manual edits to fix paper-speak hallucination, naming conventions, tag overflow, headline subtitle bugs, and benchmark renderer fallthrough — every one a manual touch on hand-authored prose. The architectural target going forward is: all the fields a forge run needs to populate an alloy MUST live as Continuum entity data inside a ForgeRecipe entity. The forge takes the recipe entity as input, runs the prune/quant/eval stages, and emits the populated alloy as OUTPUT. The forge never consumes a hand-authored alloy; the foundry generates it. Recipe entity carries the prose fields the model card renders (description, userSummary, tags, methodologyPaperUrl, limitations[]) plus the source/stages/calibration/quant tier configuration. ForgeArtifact entity is the recipe + the eval results. publish_model.py reads the ForgeArtifact, not a hand-authored alloy file. This is the next sprint after the offline-logits architecture.
joelteply
added a commit
that referenced
this pull request
Jun 3, 2026
…m introspection (task #165) Replaces service_loop's `recall_recent(8)` (newest-first by admission) with `recall_scored(now_ms, 8)` (salience × recency-decay, top-N), and emits a per-engram INFO log at the L2 → prompt seam so the cycle's recall behavior is observable, not opaque. The substrate's continual-learning property compounds through this scoring: salient + protected + recently-used engrams stay near the top; novel ones get their protection window; everything else drains toward SALIENCE_FLOOR but doesn't disappear ([[source-drain-is-the-universal- pattern]] applied at the recall layer). record_recall_hit closes the use-it-keeps-it feedback loop — without it, scoring is one-way and memory only ever decays. PR #91 (RecallMetadata sidecar) + #92 (decay tick) provided the scoring infrastructure; this slice composes them on the read path. Per Joel's 2026-06-03 "introspect all rag" directive + [[observability-is-half-the-architecture]]: every recall now emits one INFO line per delivered engram (rank, engram_id prefix, salience, content preview). Optimization can target actual scoring behavior, not guesses. Three new admission_state tests pin the contract: - recall_scored_ranks_by_salience_desc — pinned > uplifted > untouched - recall_scored_records_recall_hit_on_returned_engrams — Hebbian loop - recall_scored_respects_limit_and_empty — boundary cases Also catches up the AnalysisInput test fixtures with the model_override field added in commit 9c8a991 (4 sites in shared_analysis/mod.rs + prompt.rs). The production caller (persona/response.rs) was already updated; only the test scaffolds were behind. 19/19 admission_state tests green on Intel Mac CPU build: cargo test --lib --no-default-features \ --features 'livekit-webrtc llama/mac-cpu-only' admission_state Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Bumps @types/node from 22.16.5 to 24.10.0.
Commits
Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting
@dependabot rebase.Dependabot commands and options
You can trigger Dependabot actions by commenting on this PR:
@dependabot rebasewill rebase this PR@dependabot recreatewill recreate this PR, overwriting any edits that have been made to it@dependabot mergewill merge this PR after your CI passes on it@dependabot squash and mergewill squash and merge this PR after your CI passes on it@dependabot cancel mergewill cancel a previously requested merge and block automerging@dependabot reopenwill reopen this PR if it is closed@dependabot closewill close this PR and stop Dependabot recreating it. You can achieve the same result by closing it manually@dependabot show <dependency name> ignore conditionswill show all of the ignore conditions of the specified dependency@dependabot ignore this major versionwill close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)@dependabot ignore this minor versionwill close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)@dependabot ignore this dependencywill close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)