FlashConv
Algorithms AIMech:FlashConv DRAFT
An execution algorithm for state-space sequence convolutions that fuses FFT-based convolution in GPU on-chip memory, uses block FFT matrix multiplications and passes state between chunks for longer sequences.
Identity Notes & Concept Scope
FlashConv is the convolution-execution contribution in the H3 paper, distinct from the H3 layer and from FlashAttention's attention-computation algorithm.
Claims & Grounded Evidence
1 Claim
Mechanism
#sequence-mechanism
Fuse the FFT, pointwise spectral product and inverse FFT to avoid intermediate off-chip transfers; factor the FFT into block matrix multiplications suited to specialized hardware. When the sequence exceeds the chosen on-chip working capacity, compute chunk convolutions and propagate the state-space boundary state between chunks.
⚠️ Curated Limitations & Negative Results
- The paper's 8K on-chip threshold concerns its stated A100 setting; no device-independent cutoff or wall-clock speedup is asserted.
- Block FFT may use more arithmetic than a conventional FFT while exploiting specialized matrix hardware.
- The printed Algorithm 2 mixes chunk variable names and its exact boundary indexing was not implemented or verified here. This record supports the execution strategy, not a certified literal transcription of the pseudocode.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Convolution1DLayer
Convolution1D Layer
|
related_match | FlashConv computes one-dimensional sequence convolutions, an operation used by this AIO layer. The algorithm concerns execution and state passing rather than denoting the layer itself, so the mapping is related rather than broader-class membership. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Daniel Y. Fu, Tri Dao, Khaled K. Saab, Armin W. Thomas, Atri Rudra, Christopher Ré
|
primary paper
full_text
|
ID: arXiv:2212.14052v3
sha256:de923022ffc5…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:33:01Z
Updated: 2026-10-11T04:33:01Z
Source File:
data/curated/FlashConv.yaml