{ "schema_version": 1, "builder_plan": { "scenario": "A self-hosted assistant is too expensive because its model uses most of the GPU memory.", "change": "Try the paper's supported lower-precision weights on a copy of your serving setup.", "baseline": "The same model at its original precision, on the same GPU.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Peak memory, task pass rate, latency, and total serving cost at the same context and load.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.13031", "title": "Attention Quantization for Tabular Foundation Models", "summary": "The method replaces 16-bit attention computation in tabular foundation models with FP8-quantized queries, keys, and values executed via a custom Triton kernel with explicit FP8 matmul instructions. It costs adoption of a custom kernel, requires hardware with FP8 instruction support (Hopper-class or newer), and demands careful calibration so that the quantization error on test rows matches that on training rows. If that error alignment is not maintained, accuracy drops drastically, and the 1.7x speedup is an upper bound that depends on workload and GPU.", "rationale": "For a reader serving tabular foundation models on a single GPU, this is a targeted, low-memory inference optimization worth validating on their own models and hardware rather than adopting blindly, since it requires FP8-capable hardware, a custom kernel, and calibration to avoid severe accuracy degradation.", "technique": "FP8 attention quantization for tabular foundation models", "family": "quantizacao", "recommendation": "testar", "published": "2026-09-11", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.13031", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces 16-bit attention computation in tabular foundation models with FP8-quantized queries, keys, and values executed via a custom Triton kernel with explicit FP8 matmul instructions. It costs adoption of a custom kernel, requires hardware with FP8 instruction support (Hopper-class or newer), and demands careful calibration so that the quantization error on test rows matches that on training rows. If that error alignment is not maintained, accuracy drops drastically, and the 1.7x speedup is an upper bound that depends on workload and GPU.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.13031", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "A Triton kernel using explicit FP8 matrix multiplication on queries, keys, and values achieves up to 1.7x speedup over 16-bit kernels on TabPFN-v3 and TabICLv2, with no relevant accuracy loss across TabArena and BeyondArena.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.13031", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [ { "anchor": "S2.E1", "label": "(1)", "section": "2 Method", "role": "baseline", "latex": "\\mathrm{Attn}(Q,K,V)=PV,\\qquad\\text{with }P=\\mathrm{softmax}\\!\\left(\\alpha\\,QK^{\\top}\\right),", "mathml": "Attn⁡(Q,K,V)=P​V,with ​P=softmax⁡(α​Q​K⊤),", "context": "A softmax attention head maps queries Q∈ℝSq×d, keys K∈ℝSk​v×d, and values V∈ℝSk​v×d to", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S2.E2", "label": "(2)", "section": "2 Method", "role": "proposed_method", "latex": "s_{X}^{(h)}=\\frac{1}{448}\\max_{i,j}\\bigl|X^{(h)}_{ij}\\bigr|,\\qquad X_{8}=\\mathrm{rn}_{\\texttt{fp8}}\\!\\left(\\mathrm{clamp}\\bigl(X/s_{X},\\,-448,\\,448\\bigr)\\right),", "mathml": "sX(h)=1448​maxi,j​|Xi​j(h)|,X8=rnfp8​(clamp⁡(X/sX,−448, 448)),", "context": "We quantize to the e4m3fn FP8 format (Micikevicius et al., 2022), whose range is [−448,448]. To stay within the representable range, each quantized tensor X is represented as X^≈sX​X8 with X8 in FP8 and an FP32 dequantization scale sX. We compute the scale dynamically from the tensor itself rather than using a static calibration. For each head h we take the absolute maximum,", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-30", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }