A bounded systems/ML experiment: can a genuinely binary EAGLE drafter improve end-to-end LLM throughput enough to offset lower speculative-token acceptance?
Start with the project overview, which defines the
goal, scope, milestones, and decision criteria. See setup for
build instructions and evaluation for measurement rules.
For Codex desktop teamwork, see agent operations.
Invoke $start-goal-team with a rough objective to open the next official goal.
Requires Git, Python 3.11+, uv, and a C/C++ compiler. Run from this repository's root:
make setup
make check
make doctor
make build-cpu
./build/llama-cpu/bin/llama-server --version
make build-metal is available for Apple development machines.
make build-cuda targets SM75 on a Linux machine with NVIDIA's CUDA toolkit.
CPU and Metal runs are development checks, not evidence for the GPU hypothesis.
configs/ Versioned experiment plans (not yet a runner interface)
data/ Local prompts, calibration data, and hidden states
docs/ Project guidance, setup, and evaluation protocol
experiments/ Experiment notes and small, reviewable reports
kernels/ Future packed binary CUDA prototypes
models/ Local checkpoints and converted GGUF files
results/ Local raw runs, logs, and environment manifests
scripts/ Bootstrap checks, builds, and environment capture
src/w1a1_eagle/ Future simulation, QAT, export, and analysis code
third_party/llama.cpp/ Pinned upstream Git submodule
This initial scaffold includes the upstream runtime and build tooling. W1A1 simulation, training, kernels, model downloads, and benchmark execution are future milestones; no speedup or model compatibility has been measured yet.
Development tools are pinned in uv.lock. The llama.cpp revision is pinned by
the submodule gitlink; see third-party notes. Large
artifacts stay out of Git. Do not use git submodule update --remote for routine
setup, because it changes the experimental runtime version.
A bounded systems/ML experiment: can a genuinely binary EAGLE drafter improve end-to-end LLM throughput enough to offset lower speculative-token acceptance?
Start with the project overview, which defines the
goal, scope, milestones, and decision criteria. See setup for
build instructions and evaluation for measurement rules.
For Codex desktop teamwork, see agent operations.
Invoke $start-goal-team with a rough objective to open the next official goal.
Requires Git, Python 3.11+, uv, and a C/C++ compiler. Run from this repository's root:
make setup
make check
make doctor
make build-cpu
./build/llama-cpu/bin/llama-server --version
make build-metal is available for Apple development machines.
make build-cuda targets SM75 on a Linux machine with NVIDIA's CUDA toolkit.
CPU and Metal runs are development checks, not evidence for the GPU hypothesis.
configs/ Versioned experiment plans (not yet a runner interface)
data/ Local prompts, calibration data, and hidden states
docs/ Project guidance, setup, and evaluation protocol
experiments/ Experiment notes and small, reviewable reports
kernels/ Future packed binary CUDA prototypes
models/ Local checkpoints and converted GGUF files
results/ Local raw runs, logs, and environment manifests
scripts/ Bootstrap checks, builds, and environment capture
src/w1a1_eagle/ Future simulation, QAT, export, and analysis code
third_party/llama.cpp/ Pinned upstream Git submodule
This initial scaffold includes the upstream runtime and build tooling. W1A1 simulation, training, kernels, model downloads, and benchmark execution are future milestones; no speedup or model compatibility has been measured yet.
Development tools are pinned in uv.lock. The llama.cpp revision is pinned by
the submodule gitlink; see third-party notes. Large
artifacts stay out of Git. Do not use git submodule update --remote for routine
setup, because it changes the experimental runtime version.