One QK Channel, Many Sources: Guarding Low-Precision Attention Collapse
Authors: Shuxiao Xie, Shuyang Xie, Yuan Cao, Dezhi Ran, Wei Yang, Tao Xie
Organizations: 1Beijing Tongming Lake Information Technology Application Innovation Center (TLAIC), China · 3Harbin Institute of Technology, China · 4Key Lab of HCST (PKU), MOE; SCS, Peking University, Beijing, China · 2Fudan University Institute of Systems for Advanced Computing, China · 5Shanghai Institute of Systems for Open Computing, China
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.
KV cache quantization reduces the memory cost of long-context LLM inference, but introduces approximation error that is typically validated only empirically. Existing systems rely on average-case robustness, with no mechanism to detect or recover from failures at runtime. We present a tiered KV cache architecture that enables runtime-certified attention: INT8 keys and INT4 values are stored in GPU memory, while FP16 originals are retained in system RAM for deterministic fallback. A two-term error decomposition yields per-head, per-step bounds on (i) attention distribution distortion from key quantization and (ii) value reconstruction error. These bounds are computed online and used to drive adaptive precision selection and a multi-stage fallback ladder, which guarantees recovery to the exact dense attention output when required. Across PG-19, NIAH, and RULER benchmarks on LLaMA~3.1-8B with contexts up to 128K, the system matches dense FP16 KV quality within noise for language modelling and retrieval tasks, while recovering catastrophic failures observed in naive INT8/INT4 baselines. Value-sensitive tasks at short context expose a controlled trade-off between compression and fidelity, which can be eliminated via tighter value tolerances or FP16-value fallback. The certification is local (per-head, per-step) and does not guarantee end-to-end model correctness, but ensures that each attention computation is either bounded relative to an FP16 reference or exactly recovered via fallback. This reframes KV cache quantization as a runtime-verified computation rather than a fixed approximation. The goal is not raw speedups, but enabling safe deployment of aggressive KV compression under strict quality constraints.
Transformers commonly exhibit an attention sink: disproportionately high attention to the first position. We study this behavior in GPT-2-style models with learned query biases and absolute positional embeddings. Combining structural analysis with causal interventions, validated across natural-language, mathematical, and code inputs, we find that the sink arises from the interaction among (i) a learned query bias, (ii) the first-layer MLP transformation of the positional encoding, and (iii) structure in the key projection. Crucially, each component we identify is individually dispensable: architectures omitting each of them robustly exhibit sinks. This indicates that attention sinks may arise through distinct circuits across architectures. These findings inform mitigation of sinks, and motivate broader investigation into why sinks emerge.
Frontier large language model training consumes massive accelerator fleets and long wall-clock computation, making stability failures costly when they occur. After a numerical or a hyperparameter fault has already destabilized the training dynamics, it may continue for thousands of steps while loss and gradient norms still appear normal. We study mechanism-driven detection of training instability by deriving internal monitors from the functional role of each critical module and from the earliest computational sites where failures are expected to produce measurable signatures. For low-precision flash attention, we monitor the spectral entropy of a QK bilinear decomposition, whose first-order term becomes abnormal before the loss fully collapses. For MoE routers, we derive indicators from their role in expert selection. Our fault-injection experiments on low-precision attention, large learning-rate, and combined faults show that these signals provide distinct signatures for different failures, triggering thousands of steps before loss divergence.