{ "schema_version": 1, "builder_plan": { "scenario": "A self-hosted model uses more compute or memory than your workload can justify.", "change": "Check for a usable pruned checkpoint and runtime support before attempting model surgery.", "baseline": "The original unpruned model at the same precision, context, and load.", "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": "Task quality, actual GPU memory, serving throughput, and preparation cost.", "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": "2610.28947", "title": "Is 2:4 Sparsity Still Worth It? Semi-Structured Sparsity versus Quantization in LLM Inference", "summary": "The technique replaces dense low-precision (e.g., FP8) linear layers with 2:4 semi-structured sparsity composed with that same low precision. It costs a mandatory retraining phase to recover model quality—perplexity recovers faster than reasoning ability—and introduces sparsity metadata overhead. It fails when a lower precision tier is still available (where dropping precision is both more accurate and faster) and on Blackwell hardware, where the required retraining budget did not recover the accuracy loss from the pairwise 4:8 pattern.", "rationale": "The performance gain is specific to H100-class hardware at the lowest supported precision and requires a retraining investment, making it a conditional optimization rather than a general adoption recommendation for constrained infrastructure.", "technique": "2:4 Semi-Structured Sparsity vs. Quantization for LLM Inference", "family": "esparsidade_e_poda", "recommendation": "observar", "published": "2026-10-20", "as_of": "2026-10-24", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.28947", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The technique replaces dense low-precision (e.g., FP8) linear layers with 2:4 semi-structured sparsity composed with that same low precision. It costs a mandatory retraining phase to recover model quality—perplexity recovers faster than reasoning ability—and introduces sparsity metadata overhead. It fails when a lower precision tier is still available (where dropping precision is both more accurate and faster) and on Blackwell hardware, where the required retraining budget did not recover the accuracy loss from the pairwise 4:8 pattern.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28947", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "On an H100, a retrained 2:4 sparse model served in FP8 by vLLM achieves 27% higher prefill throughput than the dense FP8 model, with GSM8K accuracy on par.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28947", "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": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }