How gazbg/poormansllama works
This repository is a research record and tooling collection for optimizing llama.cpp on memory-constrained GPUs, not a standalone server implementation in the supplied tree. Its practical workflow is to run controlled server or kernel experiments, capture results and traces, then analyze or cross-check them. The graph preserves the major sampled research branches and the SOURCE INDEX core modules; the vendored runtime is described by the README but its implementation is not included in the supplied tree, so runtime internals are not mapped.
Main entry points: llama.cpp Vendored llama.cpp · win-arm.py Dense server arm · check-tg-predict.py Prediction verdict · server_vision.py Vision harness
Components 20
Experiment runners
- Dense server arm
tools/dense/win-arm.py - Dense planning
tools/dense_plan.py - Dense HTTP probe
tools/dense_probe.py - TG arm runner
tools/run-tg-arms.py
Trace analysis
- Expert pre-gate
tools/analyze-pregate.py - Prediction verdict
tools/check-tg-predict.py - MTP replay
tools/tg59_d1.py - Route overlap
tools/route_overlap/analyze.py - Retention cross-check
tools/route_overlap/crosscheck-retention.py - Predictor screening
tools/predictor-screen-gplr.py
Server studies
- LM review tables
tools/server_lm_review_tables.py - Canonical tables
tools/server_canon_tables.py - Server harness
tools/server_mt.py - Vision harness
tools/server_vision.py
Kernel probes
- CPU expert bench
tools/expert-cpu-bench.cpp - Sequence split probe
tools/pascal/seqsplit_probe.cpp - CPU stream bench
tools/pascal/bench_stream.cpp - CUDA MMQ check
tools/pascal/check_mmq_remainder.cpp
Other components
- Researcher
- Vendored llama.cpp
third_party/llama.cpp
Connections 17
- Researcher → Dense server arm: runs experiments
- Researcher → Server harness: runs experiments
- Researcher → Vision harness: runs experiments
- Researcher → Route overlap: analyzes traces
- Dense server arm → Dense HTTP probe: sends probes
tools/dense/win-arm.py - Vision harness → Server harness: uses harness
tools/server_vision.py - LM review tables → Canonical tables: reads tables
tools/server_lm_review_tables.py - Retention cross-check → Route overlap: reuses analyzer
tools/route_overlap/crosscheck-retention.py - Prediction verdict → MTP replay: uses replay
tools/check-tg-predict.py - CPU expert bench → Vendored llama.cpp: uses GGML APIs
tools/expert-cpu-bench.cpp - Sequence split probe → Vendored llama.cpp: uses llama APIs
tools/pascal/seqsplit_probe.cpp - CPU stream bench → Vendored llama.cpp: uses GGML APIs
tools/pascal/bench_stream.cpp - CUDA MMQ check → Vendored llama.cpp: uses CUDA APIs
tools/pascal/check_mmq_remainder.cpp - Expert pre-gate → Vendored llama.cpp: reads GGUF code
tools/analyze-pregate.py - Dense planning → Dense server arm: plans arms
- Predictor screening → Prediction verdict: screens predictors
- TG arm runner → Prediction verdict: produces traces