forge workspace

Shared development workspace for the RAMforge and forgeCORE projects.

RAMforge/   RAMforge runtime, planner, accounting, CLI/TUI
forgeCORE/  safe Rust bindings and integration layer over llama.cpp/ggml
native/     external llama.cpp/ggml checkout and build (not tracked)

RAMforge → forgeCORE → llama.cpp/ggml is the dependency direction. RAMforge must not call native llama.cpp/ggml symbols directly or depend on forge-sys.

The three Git repositories are independent:

• workspace root → NeonMC23/forge (this snapshot)
• RAMforge/ → NeonMC23/RAMforge
• forgeCORE/ → NeonMC23/forgeCORE

Nested .git/, native source/build trees, models, caches, and build outputs are excluded from this snapshot. See WORKFLOW.md for the local engineering and publication policy.

Native baseline

• llama.cpp/ggml v0.5.0
• commit 7fe450e19305b828c199d602c23a8337aaa1f03b
• shared libraries: native/build/bin
• FORGE_LLAMA_DIR selects the native tree; see forgeCORE/docs/NATIVE.md.

Current architecture coverage (2026-10-25)

Architecture	Legacy RAMforge executor	forgeCORE/libllama	Evidence
llama	yes	yes	SmolLM2-135M Q4_K_M
qwen2	yes	yes	Qwen2.5-1.5B Q4_0
gptneox / gpt_neox	not-yet (explicit, actionable rejection)	yes	Pythia-70M-deduped Q8_0; native and independent NumPy parity; GPT-NeoX-20B Q4_K_M under an 8 GiB cgroup cap
qwen3, qwen35, gemma, phi, literal mistral	not-yet	not-yet	inspect/plan only

gptneox execution must be requested explicitly with --backend forgecore; the legacy path never falls back or approximates it. The 20B GPT-NeoX GGUF (13,135,032,320 B file, 13,133,230,080 tensor bytes, SHA-256 1a35c13535c26bb925a2379d9f021de67a0390124ffe2051aa94412002652d6d) executed under a hard 8 GiB cgroup v2 cap with --enforce-ram --no-repack --no-prefetch: 7.64 GiB of the 12.23 GiB mapping resident, 7,979 limit events, no OOM kill, one greedy token Paris in 7.72 s. The same run with prefetching took 16.94 s. The opt-in no-prefetch mode avoids MAP_POPULATE via a documented native patch; it is still OS demand paging, not a RAMforge per-layer staging engine. The same patch adds model-scoped llama_model_release_pages() and RAMforge --release-mapping-pages; on the 20B prefetch run it released 7.93 GiB to 0, and decode still produced Paris. It also adds exact tensor-to-mapping metadata and --inspect-layer/--release-layer: Qwen2.5-1.5B layer 0 released 25.12 MiB with shared pages skipped, and the 20B layer 0 submitted 293.93 MiB but showed no RSS drop because those pages were already nonresident under the 8 GiB cap. A two-model fixture test released only the target mapping (A 8192->0, B 8192->8192).

RAMforge now has a graph-aware residency engine: --residency-policy off|observe|conservative|memory-bound|reuse-aware|active, a unified model-scoped memory ledger with provenance and reconciliation, in-memory execution profiles, deterministic planning, admission, bounded deduplicated staging, versioned JSON operation reports, and lifecycle guards. It releases only after the context is dropped at a safe boundary; it has no intra-decode eviction. Qwen2.5-1.5B at 1 GiB with active target 512M submitted 514.44 MiB and observed mapping RSS 855.68 -> 476.57 MiB and process RSS 1021.21 -> 642.10 MiB; the soft target was still not met (observed 642.10 MiB), and cgroup peak remained 1 GiB. Observe mode performed no mutation.

The engine also has a native execution-quiescence contract. A context holds a model-scoped shared lease from asynchronous graph submission until native synchronization/destruction; page release takes a non-blocking exclusive attempt and returns Busy while any context is in flight. forgeCORE exposes model/context identity, execution generations, quiescence, and typed residency errors. RAMforge binds those facts into its planner and resource-lifetime state machine, and only marks resources release-eligible after the operation boundary is observed quiescent. This is still between-operation physical page control, not intra-decode eviction.

RAMforge also exposes an opt-in CPU node-dispatch observer through forgeCORE. It emits one event from worker 0 before each non-empty, compute-enabled node, with model/context/execution identity, node/op/source names, bounded storage, and drop accounting. Native tensor metadata now includes model-generation- scoped mapping identity and validated page ranges, and opt-in bounded Linux mincore() sampling can run before dispatch. Dispatch and sampling are observations only: not a tensor read, page fault, I/O, completion, residency during execution, or release authorization. See RAMforge/docs/handoff/GRAPH_SAFETY.md.

RAMforge also has a deterministic predictive staging layer over validated profiles. Predictions are bounded by window/speculative budgets, admission, and confidence thresholds; they never authorize release. Feedback history is validated and atomic, and ramforge simulate provides a deterministic simulation-only comparison. A five-arm, four-workload benchmark harness with rotated order, page-fault/RSS/cgroup reporting, and an isolated-copy controlled-cache mode found no mapping/RSS reduction and no universal speedup; end-to-end timing includes prediction/staging/setup, and a 256 MiB cgroup run showed lower TTFT but worse end-to-end latency on short workloads. Linux mincore() observation now samples selected prefetch resources first, reports partial selected/general coverage separately, and labels unknown pages. A safe POSIX_FADV_DONTNEED experiment is applied only to an isolated model copy and verified with mincore; warm, controlled, and preparation-failure states stay distinct. Before-access residency remains unavailable, so eviction-before-use and causal benefit stay unknown. See RAMforge/docs/handoff/PREDICTION.md and RAMforge/docs/handoff/BENCHMARKS.md.

It also has an opt-in bounded scheduler graph trace and manifest, bounded MADV_WILLNEED staging, and a capability report. Qwen trace with a 4096-event budget retained 958 requested/958 synchronized completions, 958 nodes/1,050 resources, known graph edges, and 1,050 graph-order last uses; intra_decode_safe=false. A 128M staging request selected 95 tensors and submitted 127.72 MiB, but mapping RSS remained 0 -> 0; prefetch is a hint, not residency. On SmolLM2-135M Q4_K_M at a 128 MiB soft target (4 greedy tokens, CPU, 2 runs), memory-bound submitted 38,027,264 B and observed mapping RSS -38,027,264 B and process RSS about 163.8 -> 127.5 MiB; reuse-aware used a complete 16,384-event profile, submitted 46,702,592 B, and observed 171.8 -> 127.3 MiB. Cgroup current did not move because it is whole-cgroup accounting. Conservative selected nothing because every aligned tensor range had shared page bytes. conservative, memory-bound, reuse-aware, JSON reports, bounded/validated profile persistence (--residency-profile-load/--residency-profile-save and ramforge profile), and the benchmark harness are documented in RAMforge/docs/handoff/MEMORY_ACCOUNTING.md, RAMforge/docs/handoff/GRAPH_SAFETY.md, and RAMforge/docs/handoff/BENCHMARKS.md. The 4G/1G/100M 20B cases were not run. See RAMforge/README.md for the enforced matrix, performance, and limitations.

Current validation entry points

cd forgeCORE
FORGE_LLAMA_DIR="$PWD/../native" cargo test --workspace --locked
FORGE_GPTNEOX_MODEL=/path/to/pythia-70m.Q8_0.gguf \
  cargo test -p forge-core --test gptneox_execution --locked -- --ignored --nocapture

cd ../RAMforge
scripts/test.sh --workspace --locked
scripts/smoke-real-models.sh --no-build
RAMFORGE_GPTNEOX_MODEL=/path/to/pythia-70m.Q8_0.gguf \
  scripts/smoke-real-models.sh --no-build --include-gptneox --matrix

# enforced host-memory path (cgroup v2 delegation required)
target/release/ramforge run /path/to/qwen2.5-1.5b.gguf \
  --ram 512M --enforce-ram --ctx 256 --no-repack \
  --prompt "The capital of France is" --max-tokens 1 \
  --temperature 0 --backend forgecore --memory-report --profile
cargo test -p ramforge-runtime --test cgroup_enforcement --locked -- --ignored --nocapture
cargo test -p ramforge-cli --test enforce_ram --locked -- --ignored --nocapture

The smoke runner executes the two real models in models/ sequentially, each in a separate process, and writes commands, exit status, RSS, GPU state, model hashes, and output to a temporary evidence directory. It never modifies the model files. See RAMforge/README.md for the latest measured matrix.

Current state and limitations are recorded in forgeCORE/docs/PROJECT_STATE.md and RAMforge/docs/handoff/PROJECT_STATE.md.