Search results

perf: RoPE 3.6x slower at prefill scale due to Tensor allocation

performancePerformance optimizationPerformance optimization

Status: Open.

#389 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: Tensor::transpose 9-242x slower than ndarray at attention shapes

performancePerformance optimizationPerformance optimization

Status: Open.

#388 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: element-wise Tensor ops 1.8-2.2x slower than ndarray — consistent across add/mul_scalar

P1High priorityHigh priority

performancePerformance optimizationPerformance optimization

Status: Open.

#387 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: Q4_0/Q8_0 dequant throughput 5x below memcpy ceiling — missing SIMD vectorization

P2Medium priorityMedium priority

performancePerformance optimizationPerformance optimization

Status: Open.

#386 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: Tensor allocation overhead — systemic 2-4x penalty on memory-bound ops

P1High priorityHigh priority

performancePerformance optimizationPerformance optimization

Status: Open.

#385 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: softmax 1.6x slower than ndarray at vocab-sized dimensions

P2Medium priorityMedium priority

performancePerformance optimizationPerformance optimization

Status: Open.

#384 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: ReLU 2x slower than ndarray mapv — Tensor overhead on memory-bound op

P2Medium priorityMedium priority

performancePerformance optimizationPerformance optimization

Status: Open.

#383 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: LayerNorm 4.2x slower than ndarray scalar — same Tensor overhead as RMSNorm

P1High priorityHigh priority

performancePerformance optimizationPerformance optimization

Status: Open.

#382 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: RMSNorm 4.3x slower than ndarray scalar — Tensor allocation overhead

P1High priorityHigh priority

performancePerformance optimizationPerformance optimization

Status: Open.

#381 In paiml/aprender;

· noahgift opened on Mar 2, 2026

perf: matvec (M=1) 3.4x slower than ndarray at LLM-relevant sizes

P0Critical priorityCritical priority

performancePerformance optimizationPerformance optimization

Status: Open.

#380 In paiml/aprender;

· noahgift opened on Mar 2, 2026

feat: head-to-head criterion benchmarks for compute hotspots

Status: Open.

#379 In paiml/aprender;

· noahgift opened on Mar 2, 2026

BPE tokenizer uses O(n²) greedy rescan instead of priority queue

Status: Open.

#378 In paiml/aprender;

· noahgift opened on Mar 2, 2026