The results support intervention at the shared QK locus rather than separate repair at each fault source, and support intervention at the shared QK locus rather than separate repair at each fault source.
Abstract
A bfloat16 transformer can train normally for many steps and then collapse abruptly. Distinct low-precision errors can trigger the same failure, leaving unclear whether each source needs its own repair or one shared route can be blocked. We isolate a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where fp32 accumulation repairs it, and use the fault as an assay for moving controlled errors across sources. Errors placed outside attention still drive the same query-key (QK) spectral runaway, while correcting only QK keeps training stable with the source fault active. This source-channel dissociation shows that fault source is not failure channel. It holds across the tested architectures and scales and reproduces on a second GPU architecture. A causal probe projects each update off the current QK weights'leading three singular directions: the query projection's largest singular value stays at 11.1, whereas removing equal energy elsewhere leaves it at 237. The QK channel therefore drives the early runaway rather than merely tracking it. Entry depends on temporal sign-coherence across steps, not aggregate deviation. QK-Guard closes the channel with a dormant controller that switches on parameter-free QK normalization when attention-logit saturation begins. It contains every tested runaway and matches always-on QK normalization over 60k steps, while non-QK actions at the same trigger fail. The results support intervention at the shared QK locus rather than separate repair at each fault source.
The results suggest that bitsandbytes 4-bit quantization can impose an additional cost on applications relying on long, updatable, semantically dense contexts, even when aggregate benchmark accuracy appears largely unaffected.
Token skipping is a widely used training-free way to accelerate vision--language--action (VLA) models by bypassing computation for most visual tokens at each control step according to a gate. When the next gate is harvested from the previous accelerated forward, however, the tokens skipped at one step are also the ones least visible to the next gate, and the damage can compound across control steps until the task fails. We study the two mechanisms this class is built on, reuse and deletion, crossing each against where its gate signal comes from on identical episodes. At a skip ratio of 0.9 on LIBERO-Object, both collapse when the gate comes from the model's own accelerated forwards, to 0.68 under reuse and to 0.31 under deletion against a dense 1.00, and the collapse is invisible to the action-level detectors we evaluate. What separates collapse from dense-level operation is not the mechanism but whether the gate is clean, computed by a forward that skipped nothing. We therefore propose actuation-slack refresh, one dense pass run while the robot executes its current action chunk, off the critical path, that hands the next step a clean gate and a fresh KV base. Since the measured detectors do not reliably reveal the failure, the refresh is unconditional rather than triggered. Both mechanisms then recover to 0.98, keeping the speed of skipping and the information of a dense pass. We then integrate the refresh into state-of-the-art caching and pruning methods across two VLA policies, 4 LIBERO suites, and 4 SIMPLER tasks, where it repairs every collapse caused by using a self-harvested gate. Serve latency drops 18--22\% below dense, measured both in simulation and on a physical robot. Where the gate signal comes from, not how tokens are skipped, decides closed-loop reliability for accelerated VLAs.
Qi Luo, Shuaijun Liu, Hao Zhao et al.· 0 citations
This claim for multi-turn, tool-calling agents, where it now matters most, is tested for post-training quantization to 4-bit weights and diagnostics, the per-channel error rate and success under a shrinking budget come from logs benchmarks already collect.
Direct low-precision write-back can erase nonzero optimizer proposals. We ask what a high-precision reference trace establishes before a low-precision run. The exact target-code event is auditable coordinatewise on a realized target trajectory; pre-run aggregate projection also assumes the reference remains a useful counterfactual. In a controlled two-layer grid, 55/72 cells have measured and predicted post-initialization crossings: times span $384\times$, 52/55 are within 15\%, and 4/72 differ in category. Matched decoder experiments show stochastic rather than nearest write-back recovers most of the loss gap. A prospective analytic-grid E4M3 audit reuses one fp32 trace across three unseen NeoX-style seeds. It passes absolute-accuracy and skill gates (macro RMSE 0.00858) but fails directional specificity. In a target-outcome-blind comparison, a historical template has lower descriptive RMSE (0.00360) than the predeclared source predictor (0.00438); a post-outcome decomposition assigns 99.65\% of variation to common time, while a privileged matched-reference correction reaches 0.00283. Persistent-native Study~1 pairs three seeds across two schedules. Five cells are canonical; a manual sixth lacks canonical process identity, so the registered result remains inconclusive. A retrospective protocol-deviation analysis is negative because the complete constant-mid cohort is disjoint from the recovered cosine-restart cell. Study~2 reports mean full-SR/dead-zone-SR recoveries of 0.9766/0.9777 and a ratio of 1.0012, a policy contrast rather than causal mediation. Simulated-INT3 Study~3 replays six checkpoints and observes a 7.3071-nat (69.71\%) validation-loss reduction in one fixed seed. Exact events and write-back effects are auditable, but aggregate forecasts can reflect shared time rather than source-specific transfer.
Quantized fine-tuning (QLoRA) saves memory but not time. It dequantizes every 4-bit weight on the fly, so it trains more slowly than fp16 LoRA. We present AQLoRA (Adaptive-Quantization LoRA), a recipe that buys part of that time back. One CPU pass over the weights sets everything, with no search and no calibration data. The pass ranks layers by NF4 reconstruction error and keeps the top-K in fp16 under a memory budget. Those layers skip dequantization, which is where the speed comes from. A quality setting adapts every layer. A speed setting adapts only the top blocks, so the backward pass stops early. The rule reproduces Unsloth's hand-curated dynamic-4bit selection exactly, in seconds, where search-based allocation needs repeated calibration passes. We evaluate on Commonsense-170K across six models and four architecture families, from 1.4B to 14B. The speed setting trains 11.1 +/- 2.7% faster than well-tuned QLoRA and gives up about one accuracy point. It was faster in all nine independent timing sessions, at worst by 7%. The quality setting trains 4.8 +/- 2.4% faster. Its accuracy is level with QLoRA on every model and within a point of fp16 LoRA, for 0.2 GiB more memory. These error bars are measured between independent sessions, not within one. Earning them taught us three rules for timing on shared hardware. Fix the measurement duration, not the step count. Measure the noise floor from a duplicated arm, not a nearly identical method. Repeat whole sessions: a floor computed inside one sweep understates the real uncertainty several times over, and the random seed controls almost none of it. We validate the recipe with controls and report the two that failed. Choosing adapter layers by weight density is no better than random. Choosing protected layers by quantization error is not either. The count of protected layers, not their identity, carries the speed effect.
Practitioners accept a compressed language model once it clears a stack of data-cheap quality guards: perplexity within a small factor of the original, downstream accuracy (for example MMLU) inside a confidence interval, and data-free output-fidelity signals that compare the compressed and original network's internal representations under random probe inputs. This stack has a blind spot. Across three model families, gently-compressed models clear every guard and then invent procedure steps that were never in the instructions when they run a standard operating procedure (SOP) as an agent. The effect is operator-specific: coherent low-rank (SVD) truncation induces it, and magnitude pruning matched to the same perplexity does not. One dissociation isolates the cause. The same compressed weights that CI-win a paired output-fidelity test CI-fail the invented-step canary. The governing axis is the coherence of the compression error times its rate; the magnitude of the damage does not predict it. The data-free fidelity probe is a fidelity oracle by construction, so it cannot see this axis. We characterize the blindspot and dissociation with paired confidence intervals on a pre-registered, powered canary across three architectures. Operator-specificity replicates on all three, and the perplexity-guard evasion appears where the model admits in-guard low-rank headroom. We then give a data-free screen: a two-axis statistic of the compression error (coherent-fraction and error-rate) that flags the failing builds with fixed thresholds across architectures and matches the coherence-times-rate mechanism. Perplexity, MMLU, and fidelity acceptance do not certify agent safety. Screen gently-compressed low-rank builds before agentic deployment
I. Kennedy, T.Kennedy· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
