Search publications
“LoRA-SVD parameterizes a training-time weight as W = Wres + βL2L1. It preserves a compact principal subspace and its activation path in BF16 while computing the full-shape residual in NVFP4.” · Show in paper
“Theorem 11.1 (Realization of the estimator interfaces). Fix the construction constants and an admissible top budget Mtop = O(B). For every actual task tuple reachable from it, the row, vector and prefix interfaces defined in the preceding sections admit deterministic uniform implementations by restoring queries and cursor motions. ... 3. For a root numerical request at clock t, all records retained by the call and its descendants have total length O(Mtop + t + 1), and every descendant clock is O(Mtop + t + 1); the shared facilities use O(B + Mtop + t + 1) space. In particular an actual top request with t = O(B) uses total deterministic O(B) space.” · Show in paper
The telescoping potential argument, regenerable record design, and contour-based inverse traversal are rigorously proved.
“Specifically, GPTQ can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing the bitwidth down to 3 or 4 bits per weight, with negligible accuracy degradation relative to the uncompressed baseline. Our method more than doubles the compression gains relative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU for generative inference. ... We show experimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3.25x when using high-end GPUs (NVIDIA A100) and 4.5x when using more cost-effective ones (NVIDIA A6000).” · Show in paper
The experimental results in Tables 2, 5, and 6 directly support the 4-hour quantization time, 3-4 bit accuracy preservation, single-GPU execution on A100, and 3.24x/4.53x speedups.
“KAIROS is implemented on top of vLLM v0.18.1 [8] which supports disaggregated prefill/decode through a KV transfer API. KAIROS adds approximately 2000 lines of new code across the profiling for step latency models, the Deflection Dispatcher (which contains the analytical models and the deflection algorithm), the modifications to the scheduler on the decode nodes and plugins to report node state to the Deflection Dispatcher.” · Show in paper
“A key is a grid-valued function on the bit cube, specified by its ordered list of essential bit names and its truth table.” · Show in paper
“At each stage we replace the old description instead of carrying its history into the new message.” · Show in paper
“Table 1 lists all evaluated configurations; each changes one variable relative to the LoRA r=8, {q,v}, FP16 reference.” · Show in paper
“For a positive absolute σ < 1, to be chosen below, set bi = σvi, ηi = 1 / (1 + √(1 − ∥bi∥2 2)), Ci = I − ηibibT i .” · Show in paper
“For X \sim \mu, form the Gaussian observation h_p = B_p X + \beta_p, Cov(\beta_p, \beta_q) = . (3.1)” · Show in paper
“Lemma 7.4 (A regular-cell likelihood cutoff). There are absolute a, c1 > 0 and a dimension threshold such that the following holds. Let P be regular in a cell C, let A ∼ γn,d be independent of z ∼ P, and let W = f(A, As(z)) for a Borel map f with at most N values.” · Show in paper
“The local orbit includes the higher active determining slots, so its replacement must retain the formal identity π^0 g = PY from Lemma 7.5. This identity will restore the old integer arguments exactly in Section 9.” · Show in paper
“Retaining the completed target while reversing its auxiliary computation follows the reversible-cleanup principle of Bennett [4, pp. 525–526]. The two-buffer schedule, local transitions, and certification guarantees needed here are established in the following proof.” · Show in paper
“Choose an integer m and independent uniform binary words of length m, one for each letter, and put u = floor(m/8). We seek an absolute delta > 0 such that every pair of length-u intervals from differently labelled codewords has longest common subsequence at most (1 - delta)u.” · Show in paper
“The first stage trains only ψ, with both anchors frozen, and combines two losses: an alignment loss that keeps the stitched cache close to the native cache of Md, and a distillation loss that makes Md, reading the stitched cache, match the next-token distribution of Mp.” · Show in paper
“Y = (Xdiag(s) −1 ) · (diag(s)W) = XˆWˆ (3)” · Show in paper
“Our extensive experiments show that QQQ achieves performance on par with existing state-of-the-art LLM quantization methods while significantly accelerating inference, achieving speed boosts up to 2.24 ×, 2.10×, and 1.25× compared to FP16, W8A8, and W4A16, respectively.” · Show in paper
The reported perplexity, zero-shot accuracy, and vLLM throughput benchmarks on LLaMA models support the claimed parity with SOTA quantization methods and the specific speedup figures.
“Comparing full fine-tuning (FFT) with quantized low-rank adaptation (QLoRA) at ranks 8, 16, 32, and 64, we find that rank induces a clear acquisition–retention frontier. Low-rank QLoRA preserves out-of-domain (OOD) performance but acquires fewer facts, whereas higher ranks improve same-fact paraphrase generalization at an increasing cost in performance on unrelated benchmarks. FFT behaves as a conservative baseline: it retains general capabilities well, but does not reach the highest factual-acquisition regime.” · Show in paper
The five-seed results in Figure 1 and Table 5 directly support the full claim, including the relative positions of FFT and QLoRA ranks 8, 16, 32, and 64.
“alpha = log 3, alpha_m = floor(log(3^m))/m, e = floor(g(1/2 + alpha_m(k - 1/4))).” · Show in paper
“For a nonraw row v of width r, minimize over all original popular centres (D, r0) in the block the cost b + 10a + 10c+ + (−c)+, b = − log(Dr0), a = log v (v, D) , c = log r r0 , (3.3)” · Show in paper
“For all sufficiently large d, set k = 2\lfloor d/16\rfloor, m = k, t = m + 1, \ell = 4k. Suppose p = f\sigma_d is a probability with 0 \le f \le L < \infty. In the block experiment above, let A have k rows and let B have \ell rows. There is an absolute constant C such that I(S; W \mid B, BS) \le H(W)/t + Cd + C\log(2 + \log L).” · Show in paper
All integrability conditions, null-set transfers, and cancellations are rigorously verified in Sections 3 and 4.