How dhnihaoya/evspark works
EvSpark is a Python library and research workflow for lossless speculative DNA generation with Evo2. The main user path validates a DNA prompt, loads an Evo2 target and drafter checkpoint, drafts token blocks, verifies them against target probabilities, and returns generated sequence and metrics; README documents greedy identity with native decoding. Supporting branches cover drafter-data preparation and distillation, plus a regulatory-DNA design workflow. State-slice rollback supports hybrid-model verification. Some subsystem links rely on the README or SOURCE INDEX references; unsampled implementation details are not inferred.
Main entry points: run.py Design runner · api.py Public API · loop.py Speculative loop · data.py Training windows
Components 24
Generation runtime
- Public API
evspark/api.py - Checkpoint resolver
evspark/checkpoints.py - Speculative loop
evspark/specdec/block/loop.py - Neural drafter adapter
evspark/specdec/block/neural_draft.py - Block verifier
evspark/specdec/verifier.py - Block forward driver
evspark/specdec/block/driver.py - State slice rollback
evspark/specdec/block/slice.py - Generated sequence
evspark/api.py
Drafting methods
- Markov drafting
evspark/specdec/markov.py - Prompt lookup
evspark/specdec/lookup.py
Drafter training
- Distilled drafter model
evspark/train/drafter.py - Genome parsing
evspark/specdec/genome.py - Training windows
evspark/train/data.py - Training corpora
evspark/train/corpora.py - Online distillation
evspark/train/distill.py
Regulatory design
- Design preparation
scripts/regulatory_design/prepare.py - Design protocol
scripts/regulatory_design/protocol.py - Design runner
scripts/regulatory_design/run.py - Sequence generation
scripts/regulatory_design/generation.py - Candidate scoring
scripts/regulatory_design/scoring.py - Design reporting
scripts/regulatory_design/report.py
Other components
- Library user
- Model checkpoint hubs
- Evo2 target
Connections 28
- Library user → Public API: calls
- Public API → Checkpoint resolver: resolves checkpoint
evspark/api.py - Checkpoint resolver → Model checkpoint hubs: downloads from
evspark/checkpoints.py - Public API → Evo2 target: loads Evo2
evspark/api.py - Public API → Neural drafter adapter: loads drafter
evspark/api.py - Public API → Speculative loop: starts generation
evspark/api.py - Speculative loop → Neural drafter adapter: requests proposals
evspark/specdec/block/loop.py - Speculative loop → Block forward driver: verifies target block
evspark/specdec/block/loop.py - Speculative loop → Block verifier: checks proposals
evspark/specdec/block/loop.py - Block forward driver → Evo2 target: runs block forward
- Block forward driver → State slice rollback: uses state slices
evspark/specdec/block/driver.py - Neural drafter adapter → Distilled drafter model: adapts model
evspark/specdec/block/neural_draft.py - Speculative loop → Generated sequence: returns tokens
evspark/api.py - Library user → Markov drafting: can use
- Library user → Prompt lookup: can use
- Training corpora → Training windows: uses loaders
evspark/train/corpora.py - Training windows → Genome parsing: cleans sequences
evspark/train/data.py - Training windows → Markov drafting: uses transitions
evspark/train/data.py - Online distillation → Training windows: consumes windows
evspark/train/distill.py - Online distillation → Distilled drafter model: trains drafter
evspark/train/distill.py - Online distillation → Markov drafting: evaluates baseline
evspark/train/distill.py - Library user → Design runner: runs design workflow
- Design runner → Design preparation: calls preparation
scripts/regulatory_design/run.py - Design runner → Design protocol: uses protocol
scripts/regulatory_design/run.py - Design runner → Sequence generation: calls generation
scripts/regulatory_design/run.py - Design runner → Candidate scoring: calls scoring
scripts/regulatory_design/run.py - Design reporting → Design runner: reports results
scripts/regulatory_design/report.py - Design reporting → Design protocol: uses protocol
scripts/regulatory_design/report.py