AMD Developer Hackathon · ACT II — Track 1
Hybrid Token-Efficient
Routing Agent
Complete every task with the fewest Fireworks tokens possible —
by answering locally instead of calling the cloud.
● local_only = 0 Fireworks tokens
Dockerized · linux/amd64
Local Qwen2.5-3B + deterministic solvers
ghcr.io/asdsyd/amd-track1:v1-local-only · github.com/asdsyd/amd-track1
The challenge
Two axes decide the leaderboard
AXIS 1 — GATE
Output accuracy
Answers must stay above the accuracy gate. Local-model answers count fully.
AXIS 2 — RANK
Fireworks tokens
Ranked ascending — fewer is better. Only inference through FIREWORKS_BASE_URL counts.
Raw compute is free. Routing intelligence wins — the cheapest path that still clears the gate.
Track 1 · Hybrid Token-Efficient Routing Agent
02
Our strategy
Answer locally, spend nothing
✓
Local models count fully toward accuracy but cost zero Fireworks tokens.
✓
So we run everything in-container: deterministic solvers first, then a bundled Qwen2.5-3B.
✓
local_only mode is the floor of the token axis — the game becomes clearing the gate.
0
tokens is unbeatable
A hybrid fallback exists — but only escalates genuinely uncertain tasks in one batched call.
Track 1 · Hybrid Token-Efficient Routing Agent
03
How it works
A cost ladder, cheapest rung first
→
ROUTER · 0 tokens
Classify
8 categories
→
SOLVE · 0 tokens
Deterministic tools
+ local Qwen2.5-3B (GGUF · llama.cpp)
→
VALIDATE
Sanity-check
answer shape
→
OUTPUT
/output/
results.json
Hybrid mode (optional): only genuinely low-confidence tasks are escalated — batched into one Fireworks call, targeting < 1,500 tokens. In local_only, nothing leaves the box.
Track 1 · Hybrid Token-Efficient Routing Agent
04
Coverage
One purpose-built solver per task type
Math
safe-AST eval · 0 tokens
Code debug
sandboxed execution
Sentiment
lexicon + local LLM
Summarization
constraint-checked
Green = provable at zero cost · Blue = local model, still zero Fireworks tokens. Deterministic-first; the cloud is never touched in local_only.
Track 1 · Hybrid Token-Efficient Routing Agent
05
Results
Clears the gate at the token floor
0
Fireworks tokens
(local_only)
8/8
practice tasks correct
in-container
~87.5%
56-task local eval
accuracy
Runs comfortably in budget: full task set finishes well under the 10-minute cap on 2 vCPU / 4 GB — a 540 s internal deadline shrinks per-task budgets as time runs low.
Track 1 · Hybrid Token-Efficient Routing Agent
06
Engineering
Built to always return a valid answer
Crash-safe by construction
A valid answer is pre-seeded for every task and flushed atomically after each one — a kill or timeout never loses completed work.
Graceful degradation
If the local model can't load, deterministic solvers carry the run. The container still exits 0 with valid JSON.
Portable amd64 image
Multi-stage build, model baked in. llama.cpp is compiled from source (glibc, AVX2) so it loads on the judging hardware.
Deadline manager
A 540 s internal budget keeps the run under the 10-minute cap and writes results incrementally throughout.
Track 1 · Hybrid Token-Efficient Routing Agent
07
Run it
Pull the public image and go
# grader-style run — 4 GB / 2 vCPU, mounted I/O
docker run --rm --platform linux/amd64 --memory 4g --cpus 2 \
-v "$PWD/input:/input:ro" -v "$PWD/output:/output" \
ghcr.io/asdsyd/amd-track1:v1-local-only
Image ghcr.io/asdsyd/amd-track1:v1-local-only
Repo github.com/asdsyd/amd-track1
public · linux/amd64 · ~2 GB
Writes a valid results.json for every task, exits 0, and spends
0 Fireworks tokens — the floor of the token axis.
AMD Developer Hackathon · ACT II — Track 1
08