Skip to main content

Module attention_eval

Module attention_eval 

Source
Expand description

Attention evaluator for the optimizer loop.

Structs§

AttentionOptResult
Result of optimizing Attention for a single problem size.
AttentionProblemConfig
A single Attention problem size with its FlashAttention v2 baseline time.
RealAttentionEvaluator
Evaluates Attention kernel parameters by modeling execution time based on actual kernel characteristics and computing efficiency vs FlashAttention v2.

Functions§

generate_attention_milestone_json
generate_attention_milestone_report
optimize_all_attention_problems
optimize_attention_problem
standard_attention_problems
Standard Attention problem configurations with FlashAttention v2 baselines.