Chimeraforge

A local-first, model-agnostic LLM deployment planner. It turns "which model, quantization, GPU, and backend -- how many, will it fit, will it hit my SLO, what will it cost" into a fast, honest, measured answer, from your shell, your Python, or your AI assistant.

uvx chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB"

The trust principle

Every number is labeled measured, extrapolated, derived, estimated, or unknown, and the tool refuses to fake the ones it can't stand behind. VRAM and KV-cache are derived -- exact arithmetic over the model's real architecture, not a measurement. Throughput is a measured lookup only on the rig the corpus was measured on; on any other GPU that row is scaled by memory bandwidth and reported as extrapolated, carrying the row it came from, the rig it was measured on and the ratio applied, because a 17.8x bandwidth extrapolation (RTX 4080 Laptop 432 GB/s -> B200 7700 GB/s) is not a measurement of your card. Failing that it is an explicit roofline estimate -- never presented as data it isn't. Quality below the bundled corpus reports unknown, not a made-up score. A 0-result plan names the exact gate that rejected every candidate instead of a generic "nothing found." No telemetry, no phone-home, works air-gapped.

Give it a model -- a size class, a Hugging Face repo, an Ollama tag, or manual overrides for an unreleased model -- and it searches the (model x quantization x backend x GPU count x tensor/pipeline parallelism) space against VRAM, quality, latency, cost, energy, and an opt-in safety gate, then hands back the cheapest config that meets your SLO.

14 commands, one tool: plan - suggest - measure - workload - validate - doctor - catalog - safety - bench - eval - compare - refit - report - mcp.

The empirical corpus traces to Technical Reports TR108-TR137 (~204,000 real measurements on consumer GPUs). See the CHANGELOG for the full feature history.

Install

Try it with no install:

uvx chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB"
pipx run chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB"

Install for real:

pip install chimeraforge            # planner + model resolution (HF/Ollama) + suggest/measure/safety/bench
pip install "chimeraforge[bench]"     # + GPU environment metadata for benchmarks (pynvml)
pip install "chimeraforge[mcp]"       # + MCP server so Claude/GPT/Cursor can call the planner
pip install "chimeraforge[eval]"      # + quality evaluation (ROUGE-L; BERTScore additionally needs `bert-score` + torch)
pip install "chimeraforge[refit]"     # + coefficient refitting (numpy, scipy)
pip install "chimeraforge[all]"       # everything

Python 3.10+. The core install covers the planner and network-facing commands (httpx is a core dep). plan / suggest / catalog run fully offline; bench / measure / safety need a running backend (Ollama, vLLM, TGI, or SGLang; safety supports Ollama only). Windows / macOS / Linux.

Quickstart

# Plan a registry size class on your GPU
chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB" --request-rate 2.0

# Plan ANY model -- a Hugging Face repo or an Ollama tag
chimeraforge plan --model Qwen/Qwen2.5-7B-Instruct --hardware "RTX 4090 24GB"
chimeraforge plan --model ollama:qwen3:14b --ollama-url http://localhost:11434

# Split a model too big for one GPU across several (tensor parallelism)
chimeraforge plan --model Qwen/Qwen2.5-72B-Instruct --hardware "H100 80GB" --tp 4

# Shrink the KV-cache, print the cost/latency/quality trade-off menu
chimeraforge plan --model-size 8b --hardware "RTX 4080 12GB" --kv-quant q8 --pareto

# Benchmark a live model and plan on the MEASURED numbers
chimeraforge plan --model qwen3:14b --measure

# Discover + rank what fits your GPU and budget
chimeraforge suggest --source ollama --hardware "RTX 4090 24GB" --budget 500

Plan with your traffic, not your guesses

chimeraforge workload --from-log requests.jsonl --out workload.json
chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB" --workload-profile workload.json

Derives the request rate, prompt and output lengths, traffic variance and prefix-cache hit rate from a request log or a live vLLM/SGLang /metrics endpoint. The variance one matters most: plan otherwise takes it as one of four presets, and it drives the whole queueing tail.

Metric names are per-engine and explicit -- vLLM has renamed two of these between versions, and a scraper that silently falls back to a stale name reports a fabricated measurement. An unknown engine is an error, and a field the source did not expose stays absent rather than acquiring a default.

Decision briefs

chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB" --request-rate 2 --report brief.md

Writes a markdown record of the decision: the recommendation, every assumption as an input rather than a finding, the alternatives table, the planner's warnings verbatim, and the exact command that regenerates it. Each number is tagged measured / extrapolated / derived / estimated / unknown in prose, not just with a symbol.

It refuses to render on a stale price snapshot and exits non-zero, rather than printing an old price in a nicer font -- a formatted document reads as more durable than a terminal line, and its reader will not re-derive the arithmetic.

MCP server -- give Claude / GPT / Cursor the same numbers

GPU sizing is exactly where assistants fail: training-cutoff hardware prices and specs, plus error-prone KV-cache/batching arithmetic done from memory. chimeraforge mcp runs a stdio MCP server so an assistant calls the real planner against measured data instead of guessing.

pip install "chimeraforge[mcp]"

Claude Code:

claude mcp add --transport stdio chimeraforge -- uvx --from "chimeraforge[mcp]" chimeraforge mcp

Claude Desktop / Cursor (add to your MCP config file):

{
  "mcpServers": {
    "chimeraforge": {
      "command": "uvx",
      "args": ["--from", "chimeraforge[mcp]", "chimeraforge", "mcp"]
    }
  }
}

The --from "chimeraforge[mcp]" pulls in the MCP SDK; uvx runs the server in a self-contained environment. If you have already pip install "chimeraforge[mcp]" into the environment your client launches, you can instead use "command": "chimeraforge", "args": ["mcp"].

Exposes five tools: chimeraforge_plan (the full gate search), chimeraforge_suggest (the inverse -- rank what actually fits a given GPU), chimeraforge_compare_api (self-host vs hosted-API cost and the break-even volume), chimeraforge_resolve_model (grounds a model id in its real params/architecture), and chimeraforge_list_hardware. Every result carries the same measured / extrapolated / estimated / unknown provenance as the CLI, and the tool descriptions tell the model to prefer them over its own knowledge. chimeraforge_plan also returns a launch field -- the serve command for the recommended config -- so the assistant can answer "and how do I run it" without inventing flags. chimeraforge_compare_api prices against a dated snapshot and reports its age, so an assistant quotes a price with its capture date rather than presenting a stale figure as current.

Commands

plan -- predictive capacity planner

chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB" --request-rate 2.0
chimeraforge plan --model Qwen/Qwen2.5-7B-Instruct --hardware "RTX 4090 24GB"   # any HF repo
chimeraforge plan --model ollama:qwen3:14b --ollama-url http://localhost:11434  # any Ollama tag
chimeraforge plan --model Qwen/Qwen2.5-72B-Instruct --hardware "H100 80GB" --tp 4   # multi-GPU
chimeraforge plan --model-size 3b --kv-quant q4 --pareto                       # smaller KV cache, trade-off menu
chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB" --launch          # + the serve command to actually run it
chimeraforge plan --model-size 3b --workload agent --safety-target 0.85 --json

• Plans any model: registry size class, HF repo (org/name), Ollama tag, or manual overrides (--params-b/--n-layers/...).

• Searches (model x quantization x backend x N-replicas x batch/GPU) through a 5-gate pipeline: VRAM -> quality -> safety (opt-in) -> latency -> budget.

• Models real serving physics: continuous batching (vLLM/TGI), prefill/decode split (TTFT + TPOT), KV-cache-bound concurrency, and variance-aware queueing (--workload).

• Fits models too big for one GPU: --tensor-parallel/--tp {N|auto} shards weights + KV across N GPUs (Megatron-style, comms-modelled); --pipeline-parallel/--pp {N|auto} splits layers across N stages instead (cheaper on slow interconnects, needs batching to fill the pipeline). Not combinable yet.

• Serves what the backend serves: GGUF quants are offered on Ollama; vLLM/TGI get FP16 and FP8 (only on GPUs with FP8 tensor cores -- Ada/Hopper/Blackwell/CDNA3). The planner no longer suggests a GGUF checkpoint on vLLM priced with a llama.cpp speedup.

• Runs where the engine runs (--platform linux|windows|wsl2|macos, default linux): each engine is offered only where its own docs say it runs on that OS and GPU. Refusals happen only on a documented statement, and each one cites the doc:

  • vLLM on native Windows;
  • SGLang or TGI on a consumer Radeon card, where their ROCm docs are Instinct-only;
  • vLLM AWQ/GPTQ on AMD, or FP8 on Intel;
  • TGI AWQ on ROCm.

  Where an engine's docs are silent the plan warns that the configuration is unverified rather than refusing it, and TGI's maintenance mode is always stated.

• KV-cache quantization (--kv-quant {fp16,q8,q4}) shrinks the cache and raises max concurrency -- biggest win at long context.

• Heterogeneous fleets (--fleet "H100 80GB,A100 80GB,L4 24GB"): sizes a mix of GPU types instead of N copies of one, because a cheap GPU can win at loose SLOs and small requests while an expensive one wins at tight SLOs and long requests. On an 8B at 250 req/s that is 3x H100 + 1x L4 at $5,760/mo against 6x A100 at $6,912 -- 16.7% cheaper, because the last few req/s are cheaper on a small GPU than on another big one (plan --model-size 8b --request-rate 250 --fleet "H100 80GB,A100 80GB,L4 24GB" --budget 100000). A mix presumes a capability-aware router that no serving engine ships, so every mixed plan says so, and the reported provenance is the worst across the types used rather than the best.

• Cost realism (--duty-cycle, --gpu-price-multiplier): the headline $/1M-tok prices a saturated fleet. You also pay for provisioned headroom and for every idle hour, so the effective figure on an 8B at 2 req/s on an H100 is $2.71/1M at full duty and $9.04/1M at 30%, against $0.71 at capacity (plan --model-size 8b --request-rate 2 --hardware "H100 80GB" --budget 100000 --duty-cycle 0.3). Spot/reserved pricing is your input, not a bundled guess.

• Self-host vs API break-even (--compare-api): prices your workload against hosted APIs and reports the monthly volume where self-hosting starts winning. Prices are a dated snapshot with a source URL per provider, flagged stale past 90 days -- never presented as a live quote -- and a frontier API is labeled as a different quality tier rather than passed off as like-for-like.

• Prefix caching (--prefix-cache-hit-rate): chatbot and agent traffic reuse a long system prompt, so most of the prefill is already cached. At a 4k prompt and a 90% hit rate an 8B on an H100 goes from 166ms to 17ms TTFT (plan --model-size 8b --prompt-tokens 4096 --hardware "H100 80GB" --budget 100000 --prefix-cache-hit-rate 0.9); the same query on the reference RTX 4080 is 2051ms to 205ms. Defaults to 0 and is never inferred, and the KV a shared prefix saves is deliberately not deducted -- under-sizing KV is what turns "it fits" into an OOM.

• Reasoning models (--reasoning-tokens N): hidden thinking tokens are decoded by the GPU and held in KV even though the caller never sees them. Counting only visible output under-counts decode by the reasoning ratio -- 1000 hidden tokens take an 8B plan on an H100 from 193ms to 3664ms p95 (plan --model-size 8b --hardware "H100 80GB" --budget 100000 --reasoning-tokens 1000). Defaults to 0 and is never inferred: the ratio is a property of your workload, not the weights.

• Attention-shape aware KV: MLA (DeepSeek-V2/V3) caches a compressed latent rather than per-head K/V -- sizing it as GQA overstates DeepSeek-V3's cache by 57x -- and sliding-window models stop growing the cache past the window. A window whose layer pattern isn't declared is not applied, because under-sizing KV is what turns "it fits" into an OOM.

• Mixture-of-Experts aware: VRAM sizes on total params (every expert stays resident) while throughput and TTFT use active params (a token only reads the experts it routes to). Treating an MoE model as dense under-predicts its throughput by 3.6x on Mixtral-8x7B and ~18x on DeepSeek-V3. Active counts are derived from the model's real expert geometry and match published figures.

• Energy (--electricity-rate): monthly kWh cost, $/1M-tok (+energy), and tok/s-per-watt, reported alongside (not folded into) the budget gate.

• Launch-command export (--launch): emits the vllm serve / ollama run / TGI docker run command for the winning config, with the plan's own context length, TP/PP degree, batch size, and KV dtype filled in -- the flags that are error-prone to hand-compute. It won't fabricate what it can't derive: a GGUF quant level becomes a note to serve the native-equivalent checkpoint, not an invented --quantization flag.

• Per-prediction provenance (measured / extrapolated / derived / estimated / unknown); explains the binding gate when nothing fits.

• Validated on registry data: VRAM R^2=0.968, throughput R^2=0.859, quality RMSE=0.062, latency MAPE=1.05% (beats analytical M/D/1 by 20.4x, TR133). No ML -- empirical lookup tables with first-principles interpolation (roofline for off-registry models).

suggest -- discover & rank models

chimeraforge suggest --source ollama --hardware "RTX 4090 24GB" --budget 500
chimeraforge suggest --source hf --hf-limit 8 --hardware "RTX 4080 12GB"
chimeraforge suggest --source catalog --hardware "RTX 4080 12GB"   # offline, after `catalog --build`

Pulls candidates from a live Ollama (/api/tags), the HF Hub (top text-generation), and/or the local catalog; resolves each to real params/arch, runs the same gate search, and shows the best config per model.

measure -- benchmark live, plan on real numbers

chimeraforge measure --model qwen3:14b --ollama-url http://localhost:11434
chimeraforge plan --model qwen3:14b --measure   # measure then plan in one step

Benchmarks the live model (real N=1 throughput, service time, concurrency scaling) and folds it into a local corpus. plan / suggest then prefer the measured numbers automatically (provenance flips to measured).

SGLang ships with no measured rows, and the planner says so rather than borrowing vLLM's. Measure your own: chimeraforge measure --backend sglang --model <served-model-name> --base-url http://localhost:30000. The SGLang adapter streams and times first and last token, so it records the decode rate the planner predicts rather than tokens over wall clock, which includes prefill. It takes the token count from the server's usage block, and a response without one is discarded, not estimated.

workload -- derive plan inputs from real traffic

chimeraforge workload --from-log requests.jsonl --out workload.json
chimeraforge workload --from-metrics http://localhost:8000/metrics --engine vllm --out workload.json
chimeraforge plan --model-size 8b --hardware "RTX 4090 24GB" --workload-profile workload.json

Reads the request rate, prompt/output lengths, traffic variance and prefix-cache hit rate off a JSONL request log or a live vLLM/SGLang /metrics endpoint, so plan stops taking them as typed-in guesses. The variance one matters most -- it drives the whole queueing tail, and a measured CV^2 is not one of four presets.

Metric names are per-engine and explicit; an unknown --engine is an error and pointing the wrong one at an endpoint fails loud, because a scraper that silently falls back to a renamed metric reports a fabricated measurement. A log yields measured mean and variance; a Prometheus histogram yields an exact mean but a bucket-approximated variance, labeled estimated. A single scrape is not a rate, so request_rate stays absent rather than being divided out of an unmeasured uptime -- and any field the source did not expose stays a required input to plan, never a default. An explicit flag always beats the profile.

validate -- audit predictions against measurements

chimeraforge validate --matrix matrix.json --measurements captured.json

Scores the planner's own predictions by provenance class, so "estimated" carries a number instead of a vibe. The config matrix is fingerprinted into the audit (SHA-256, order-independent). Pass that hash back with --expect-fingerprint <hash> and the command fails unless the matrix still hashes to it, so a matrix edited after seeing results cannot be passed off as the one that was registered -- pre-registration, not post-hoc selection. Without the flag the fingerprint is recomputed from whatever matrix was loaded and only printed, which proves nothing on its own. Every cell is published, the worst case survives aggregation rather than being averaged away, and a class with too few cells is labeled underpowered instead of quoted as a rate.

Measurements are sourced records, not bare numbers. Each cell states:

• who measured it: own-rig-measured or third-party-measured, which are separate scorecard rows and never averaged together;
• where it came from: a source_url, or for own-rig runs the recorded bench environment;
• a captured_at date;
• whether the source omitted its serving config (underspecified), which keeps the cell out of the headline rows but still publishes it;
• which quantity each number is: decode_tps_single_stream, e2e_tps_single_stream, prefill_tps (converted to TTFT at the cell's prompt length), ttft_ms, e2e_latency_ms (one request, batch 1, scored against the planner's service time), e2e_latency_p95_ms or aggregate_tps_at_concurrency.

An end-to-end or aggregate rate is kept in the raw output but never scored against the planner's decode prediction, and ambiguous does not load. A third-party cell citing the TR corpus the planner was fitted on is refused. The v1 shape, {cell: {metric: value}}, is refused as unsourced.

{"schema_version": 2, "hardware": "RTX 4090 24GB", "cells": {
  "<model|quant|backend|c..|p..|o..|b1>": {
    "evidence": "third-party-measured", "source_url": "https://...", "captured_at": "2026-09-25",
    "underspecified": false, "config_quote": "llama-bench -ngl 99 -fa 1",
    "metrics": [{"definition": "decode_tps_single_stream", "value": "<tok/s>", "quote": "tg128 | ..."}]}}}

The scorecard reports in-band pass rate (against a bands tolerance pre-registered in the matrix; n/a when none was), median absolute error, GMFE (geometric mean fold error, so 2x high and 2x low score the same), signed bias, and the worst cell. Bands, the consulted-source list and any per-cell spec (architecture pinned for offline prediction) are part of the fingerprint. Predictions always come from the bundled corpus unless --models-path is given, so a published audit does not depend on what measure left in your cache. A batched cell is skipped rather than compared to a single-stream prediction.

The published audit. corpora/SCORECARD.md grades the planner against 42 cells from 8 published third-party benchmark sources on 17 GPUs. The raw JSON, every exclusion with its rule, and the pre-registered source list are all committed, and the write-up is TR147. Every audited prediction is a roofline estimate.

On fully specified cells, decode is inside +-25% only 13% of the time, with a median absolute error of 35.9% (n=15). The errors split by memory type in opposite directions:

• HBM datacenter parts: decode is over-predicted by a median of +58%, up to +206% on a B200. llama.cpp measures 200-308 tok/s across A100/H100/H200/B200/MI300X, while the roofline scales with bandwidth.
• GDDR consumer cards: decode is under-predicted by a median of -36%.

Read a roofline decode figure on an HBM part as an upper bound. Regenerate the audit with python scripts/build_validation_corpus.py --write --audit. A test fails if the published audit goes stale or its error bands widen.

doctor -- check this machine (read-only)

chimeraforge doctor            # detected GPUs, what the planner can do with each, local engines
chimeraforge doctor --json     # the same report as JSON

This command detects the local platform with each vendor's own tool and changes nothing. Each probe names the tool it used, and a missing tool is reported as a finding, not an error.

• NVIDIA: nvidia-smi for devices, pynvml for the CUDA version.
• AMD on Linux: amd-smi, or the deprecated rocm-smi. When a card is unlisted, its own reported bandwidth fills --gpu-bandwidth-gbps.
• Apple Silicon: system_profiler, reported as unified memory.
• Intel: xpu-smi. Its PCI-ID device names are never guessed into a product.
• Windows, any vendor: CIM plus the display-class registry. Win32_VideoController.AdapterRAM is a uint32 and caps at 4 GB, so the driver's qwMemorySize is read instead. An integrated GPU's figure is labelled a shared-memory aperture, not VRAM.
• WSL: detected from WSL in the kernel release. Microsoft notes "microsoft" alone appears in non-WSL kernels.

Exit codes are not trusted, since rocm-smi exits 0 with nothing to report. The parsers are golden-tested against real captures, with each source and license listed in tests/fixtures/doctor/SOURCES.md.

Every device gets a planner status:

• matched: a database entry, plus the flag to plan it with.
• supply-figures: not in the database, so the --gpu-* flags it needs are listed. A number is filled in only where the tool reported dedicated VRAM.
• not-representable: for example unified memory, which the planner cannot model yet.

Local serving engines are probed on the same default URLs bench uses. An engine counts as running only when it identifies itself through its version endpoint. A generic web app answering /health on :8000 is reported as "answers but did not identify as vllm", not as vLLM. (check is reserved for plan drift detection.)

doctor also shows the engine-support matrix row for the platform it detected. The rows are linux-cuda, linux-rocm, linux-xpu, windows-native, windows-wsl2, macos-apple-silicon and cpu.

Each cell comes from the engine's own docs, at a pinned release: vLLM v0.30.0, SGLang v0.5.20, TGI v3.3.7 and Ollama v0.34.4. Every claim carries a verbatim quote and a /blob/<tag>/ URL, and silence is recorded as not documented, not guessed.

What the matrix says:

• vLLM does not run natively on Windows; it runs under WSL2.
• On ROCm, vLLM lists consumer Radeon cards (RDNA3/4), while SGLang and TGI are Instinct-only.
• Ollama is the one engine documenting native Windows AMD support.
• TGI's repository is archived and in maintenance mode.

scripts/build_engine_support.py rebuilds and validates the matrix. doctor warns once it is more than 90 days old.

catalog -- local model catalog

chimeraforge catalog --build         # resolve a curated seed (+ --with-ollama) and cache specs
chimeraforge catalog                 # list the cached catalog

Persists resolved specs so suggest --source catalog ranks a known-good set fully offline.

safety -- live refusal screen

chimeraforge safety --model llama3.2-3b --prompts harmful.txt --quant Q4_K_M --safety-target 0.85

Where plan --safety-target decides from bundled TR134/TR142 data, safety measures: it runs your probe prompts against a live model, classifies refusals (rule-based -- the TR134 regex baseline), reports the measured refusal rate vs the bundled gate data (expected, drift, RTSI risk tier), and exits 1 below --safety-target. You provide the prompts (--prompts, one per line) -- no attack corpus ships with the package; point it at HarmBench / AdvBench / your own set. Needs a running Ollama.

bench -- live inference benchmarking

chimeraforge bench --model llama3.2-3b --runs 5
chimeraforge bench --model llama3.2-3b --all-quants --context 512,1024,2048,4096 --json
chimeraforge bench --model llama3.2-3b --backend vllm --base-url http://localhost:8000

Three workload profiles (single / batch / server-Poisson); measures throughput, TTFT, and latency with p50/p90/p95/p99; CV-based stability warnings; JSON output.

eval -- quality evaluation

chimeraforge eval --task general_knowledge --json
chimeraforge eval --predictions preds.txt --references refs.txt --model llama3.2-3b

Metrics: exact match, ROUGE-L (LCS fallback), BERTScore, coherence -> composite (0.2*EM + 0.3*ROUGE + 0.3*BERT + 0.2*coherence). Quality tiers from TR125; 3 built-in tasks (general_knowledge, summarization, code). Pass --fp16-baseline to classify the drop tier.

compare -- diff benchmark runs

chimeraforge compare --baseline run1.json --candidate run2.json,run3.json --json

Matches configs by (model, backend, quant, workload, context_length); computes throughput/TTFT/duration deltas with an aggregate improvement/regression summary.

refit -- update planner coefficients

chimeraforge refit --bench-dir ./results/ --output fitted_models.json --validate

Bayesian blending (per-key confidence weighting), hardware offsets, power-law refitting, and a 10-check validation suite that gates the write (--validate).

report -- generate reports

chimeraforge report --results-dir ./results/ --format markdown --output report.md

Markdown (GitHub-compatible) and self-contained, XSS-safe HTML; statistical analysis (RMSE, MAE, MAPE, R^2) with per-config percentile tables.

mcp -- serve the planner to AI assistants

chimeraforge mcp

Runs the stdio MCP server described above. Requires pip install "chimeraforge[mcp]".

What's modeled

Hardware: 35 GPUs:

• Consumer: Ampere/Ada/Blackwell (RTX 30/40/50-series) and AMD RDNA4 (RX 9070 / 9070 XT / 9060 XT).
• Workstation: RTX PRO 6000 Blackwell (Workstation, Max-Q, Server), Radeon AI PRO R9700, and Intel Arc Pro B60/B65.
• Datacenter: A100 40/80GB, H100, H200, B200, L4, T4, and AMD MI300X/MI325X/MI350X/MI355X/MI455X.

Each has VRAM, bandwidth, FP16 TFLOPS, TDP and interconnect (NVLink/Infinity Fabric/PCIe). Each carries its source URL, capture date, the datasheet column its TFLOPS figure came from, and whether its $/hr is a rental rate or an amortised purchase.

A figure the vendor does not publish stays unknown, and unknown is not zero:

• No price (RTX PRO 6000, Instinct, Arc Pro, R9700): the budget gate refuses the card instead of pricing it at $0. Pass --gpu-price-per-hour.
• No dense FP16 figure (Arc Pro, and the RTX PRO 6000 Server Edition, whose "1 PFLOP" is unlabeled): TTFT is reported as the memory-bound floor, a lower bound, and a --ttft-slo is refused rather than checked against it. Regenerate and validate with scripts/build_hardware_data.py. An unlisted GPU is no longer a wall: --gpu-vram-gb and --gpu-bandwidth-gbps (plus optional --gpu-fp16-tflops / --gpu-tdp-w / --gpu-interconnect-gbps / --gpu-price-per-hour) plan any card, and --hardware auto reads the installed one.

Known limits (honest): Speculative decoding is not yet modeled. The prefill floor is a bound derived from MBU_DEFAULT, which is calibrated on a single datapoint -- read it as "no faster than", not as a prediction. Chunked-prefill overhead is derived from the KV re-read the mechanism implies and then clamped at the one published ceiling (25% at a 512-token budget, Sarathi-Serve arXiv:2403.02310); the tile-quantization cliff (a 257-token budget measured ~32% slower than 256) is real, sharp, and deliberately not modeled -- the planner warns instead. Prefix caching models the prefill saving but not the KV saving (deliberately conservative). Reasoning tokens are modeled but the ratio is your input (--reasoning-tokens), never inferred. For MoE, active-vs-total params are modeled, but expert parallelism and routing load-imbalance are not. For hybrids, the attention-layer split and the Mamba-2/Mamba-1/gated-DeltaNet recurrent state are read from the model's own config and derived from shapes in transformers source; Kimi's KDA state is inferred from the DeltaNet convention and says so, and a family whose layer pattern cannot be placed (Falcon-H1, a parallel hybrid) keeps full KV on every layer rather than being guessed at. Multi-LoRA sizes adapter VRAM exactly, but its decode cost is a rank-indexed estimate from a single published sweep, and per-adapter KV fragmentation is not modeled. Heterogeneous fleets solve the allocation exactly but assume a request router that no engine currently provides, and inherit the throughput-estimate error of every GPU type in the mix. Quant coverage for vLLM/TGI/SGLang is FP16 + FP8 + AWQ/GPTQ; FP8 and W4A16 quality are estimated, not measured -- the TR quality corpus only covers GGUF k-quants. The bundled quality corpus is 20 items, which resolves nothing smaller than ~21 percentage points (Miller, arXiv:2411.00640 Eq. 9), so every measured quant delta in it is reported as indistinguishable from its FP16 baseline rather than as a difference -- run a real harness and pass it with --quality-from to get a cell that can support one. Quality is measured at 2K context and reported UNKNOWN for narrow quants at >=64K, where published losses reach 59% (arXiv:2505.20276). TP and PP throughput are comms-modelled estimates, not measured, and can't be combined in one plan. Queueing is analytical (variance-aware), not a discrete-event simulator. The bundled corpus is fit primarily on one rig (RTX 4080 12GB); other GPUs scale from bandwidth/compute until you measure on yours. How far that scaling misses is now published rather than assumed: against third-party benchmarks the roofline is optimistic on HBM parts (median +58% decode) and pessimistic on GDDR cards (median -36%), and its TTFT errors run the other way (TR147). Unified-memory devices (Apple Silicon, Strix Halo, DGX Spark) cannot be represented yet; that needs its own pass rather than a guessed spec. Engine availability per platform is enforced from each engine's own docs, but the deployment OS is an input (--platform, default linux), not something the planner can see. Where an engine's docs are silent on a platform the plan warns rather than refuses. A user-supplied card has no known vendor, so its engine support is not checked, and every candidate says so. --hardware auto still reads NVIDIA only. The MCP server is stdio-only (Claude Code/Desktop, local Cursor) -- no hosted remote transport yet.

What the research decided

Phase 2 (TR123-TR133, ~106,000 measurements) distilled into an artifact-backed deployment framework -- the same rules the planner applies:

Headline findings (full data in the TRs): Rust beats Python single-agent (+15.2% throughput, -58% TTFT, -67% memory -- TR112); dual Ollama reaches near-perfect multi-agent parallelism (~99%) vs 82.2% on one instance (TR110/TR113/TR114); vLLM's continuous batching gives a 2.25x edge at N=8, bottlenecked on GPU memory bandwidth, not the stack (TR130-TR132).

Full research: docs/archive/technical_reports.md indexes all 32 reports; the full archive with methodology and raw-data references lives in outputs/publish_ready/reports/.

How the numbers are made

• ~204,000 primary measurements across 32 technical reports (TR108-TR137 + the TR142/TR146 safety provenance), on an RTX 4080 Laptop (12 GB; 192-bit GDDR6, 432 GB/s), which is the reference rig every cross-GPU estimate is scaled from. De-duplicated: TR137/TR142 are syntheses of already-counted data.
• Rigor: fresh-process isolation per run (no warm-cache bias), forced cold starts, 3-5 runs per config for statistical confidence, structured JSON/CSV logging with full provenance. Every claim traces to raw data you can re-run.
• Program context: ChimeraForge is the actionable CLI splice of the parent Banterhearts program (~1,337,000 primary + judge measurements across 54 TRs); the safety attack-surface and serving-stack research lives in sibling repos.
• 2,466 automated tests (pytest tests/) cover the planner models, gate search, resolver, discovery, safety, bench backends, and the MCP server -- GPU-decoupled, no live backend required for the core suite.

Reproduce any number: find the claim in a report under outputs/publish_ready/reports/, follow its reference to the data folder, inspect the CSV/JSON, and re-run the provided scripts or notebooks. See docs/archive/methodology.md.

Repository layout

Documentation

• docs/README.md -- documentation index and navigation
• docs/quick_start.md -- first benchmark run (Python + Rust)
• docs/API.md -- Python API reference for the package
• docs/archive/technical_reports.md -- index of all 32 technical reports
• docs/archive/dual_ollama_setup.md -- required for reproducing multi-agent results
• docs/archive/methodology.md / docs/archive/rust_vs_python.md -- methodology and the full language comparison

Contributing

Contributions welcome -- see CONTRIBUTING.md. Good areas: additional benchmark configs, new optimization strategies, more models/hardware, docs, and analysis tools.

License

MIT -- see LICENSE.

Acknowledgments

Conducted as part of the Banterhearts LLM Performance Research Program: Phase 1 (TR108-TR122) established the measurement methodology and cross-language comparison, Phase 2 (TR123-TR133) produced the deployment framework and capacity planner, and Phase 3 (TR134-TR137) measured the safety cost of inference optimization -- now the planner's opt-in safety gate.

Repository: https://github.com/Sahil170595/Chimeraforge - PyPI: https://pypi.org/project/chimeraforge/ - Status: Beta, actively developed