@atomic-chat/core 0.13.6 → 0.17.0

package/dist/runtime/tensorrt-llm/quant-format.js

@@ -1,189 +0,0 @@
-/**
- * The quantization-format naming rule: a fixed checkpoint-metadata → format-name mapping that
- * `RuntimeDescriptor.quantization[].format` (`src/contracts/environment.ts`) is keyed by. Ported
- * verbatim from `atomic-chat-conf`'s README, "Runtime descriptors (`runtimes/`)" →
- * `quantization[].format` names — this file *is* the core side of that conf↔core contract, so any
- * change to the rule has to change both places together.
- *
- * The inputs are a checkpoint's `config.json` and, when the repository carries the file, its
- * `hf_quant_config.json`. A checkpoint can be quantized with no `quantization_config` in
- * `config.json` at all (e.g. `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8`,
- * `nvidia/Llama-3.3-70B-Instruct-NVFP4`), so `hf_quant_config.json` must always be checked before
- * `dtype`/`torch_dtype` is trusted. Steps apply in order, stopping at the first that matches; `null`
- * means the naming rule does not recognise the checkpoint (an unsupported `quant_method` such as
- * `awq`/`gptq`, an `fp8` checkpoint without the block size this engine loads, or a `dtype` other
- * than `bfloat16`/`float16`) — the caller rejects that as an unsupported format, never guesses.
- *
- * Pure and network-free, like every file in this module (design D12).
- */
-const FP8_PB_WO = 'fp8_pb_wo';
-const FP8_BLOCK_SCALES = 'fp8_block_scales';
-function asObject(value) {
-    return typeof value === 'object' && value !== null && !Array.isArray(value)
-        ? value
-        : undefined;
-}
-function asNonEmptyString(value) {
-    return typeof value === 'string' && value.length > 0 ? value : undefined;
-}
-/** `fp8_pb_wo` is what NVIDIA ModelOpt calls it; the engine itself calls the same format `fp8_block_scales`. */
-function renameFp8PbWo(quantAlgo) {
-    const lower = quantAlgo.toLowerCase();
-    return lower === FP8_PB_WO ? FP8_BLOCK_SCALES : lower;
-}
-/**
- * Every sibling that is a `.gguf` file. GGUF is not a case of the naming rule at all — this engine
- * never loads a GGUF checkpoint's `config.json` — it is the separate, always-reject gate
- * `compatibility.ts` applies before the naming rule runs at all (spec `tensorrt-llm-models`: "for
- * GGUF there is llama.cpp"). A listing with even one `.gguf` file counts, which also covers a
- * GGUF-only listing.
- */
-export function isGgufCheckpoint(files) {
-    return files.some((file) => file.path.toLowerCase().endsWith('.gguf'));
-}
-/**
- * Step 1: `hf_quant_config.json` present (NVIDIA ModelOpt) → its `quantization.quant_algo`,
- * lower-cased and renamed. The README rule stops here whenever the file is present at all — a
- * present-but-empty/missing `quant_algo` (e.g. a ModelOpt KV-cache-only export) is `null`
- * (unrecognised), never a reason to fall through to `config.json`/`dtype`.
- */
-function fromHfQuantConfig(hfQuantConfigJson) {
-    const quantization = asObject(hfQuantConfigJson.quantization);
-    const quantAlgo = quantization === undefined ? undefined : asNonEmptyString(quantization.quant_algo);
-    return quantAlgo === undefined ? null : renameFp8PbWo(quantAlgo);
-}
-/**
- * Step 2: `config.json`'s `quantization_config.quant_method`. Returns `null` (recognised as *not
- * loadable*, distinct from `undefined` meaning "step 2 does not apply") for every `quant_method`
- * this engine release cannot load, so the caller never falls through to reading `dtype` for a
- * checkpoint that step 2 already identified as quantized.
- */
-function fromConfigQuantizationConfig(configJson) {
-    const quantizationConfig = asObject(configJson.quantization_config);
-    if (quantizationConfig === undefined)
-        return undefined;
-    const quantMethod = quantizationConfig.quant_method;
-    if (quantMethod === 'modelopt') {
-        const quantAlgo = asNonEmptyString(quantizationConfig.quant_algo);
-        return quantAlgo === undefined ? null : renameFp8PbWo(quantAlgo);
-    }
-    if (quantMethod === 'fp8') {
-        const blockSize = quantizationConfig.weight_block_size;
-        const isFp8BlockScales = Array.isArray(blockSize) && blockSize.length === 2 && blockSize[0] === 128 && blockSize[1] === 128;
-        // `fp8` without that exact block size is not loadable by this engine release
-        // (`model_config.py:323`) and must not fall through to step 3 as `bf16`/`fp16`.
-        return isFp8BlockScales ? FP8_BLOCK_SCALES : null;
-    }
-    if (quantMethod === 'mxfp4')
-        return 'mxfp4';
-    // awq, gptq, anything else: not loadable by this engine release.
-    return null;
-}
-/**
- * The checkpoint's weight dtype: `dtype` is what TensorRT-LLM itself reads, `torch_dtype` is the
- * legacy fallback. A VLM-style config (Qwen3.5 and later) declares it only on `text_config`, which
- * TensorRT-LLM 1.3 falls back to as well (`model_config.py`), so this does too.
- */
-function configDtype(configJson) {
-    const textConfig = asObject(configJson.text_config);
-    return (asNonEmptyString(configJson.dtype) ??
-        asNonEmptyString(configJson.torch_dtype) ??
-        (textConfig === undefined
-            ? undefined
-            : (asNonEmptyString(textConfig.dtype) ?? asNonEmptyString(textConfig.torch_dtype))));
-}
-/** Step 3: unquantized, named by `configDtype`. */
-function fromDtype(configJson) {
-    const dtype = configDtype(configJson);
-    if (dtype === 'bfloat16')
-        return 'bf16';
-    if (dtype === 'float16')
-        return 'fp16';
-    return null;
-}
-/**
- * The naming rule, applied in order. `null` means unrecognised: the caller must reject it as an
- * unsupported format rather than reporting `bf16`/`fp16` by default.
- *
- * `hf_quant_config.json` being present at all stops the rule at step 1 — its own `null` (an
- * unrecognised or missing `quant_algo`) is final and never falls through to `config.json`/`dtype`,
- * matching the README: "if hf_quant_config.json is present ... the format is its quant_algo".
- */
-export function quantizationFormat(configJson, hfQuantConfigJson) {
-    if (hfQuantConfigJson !== null) {
-        return fromHfQuantConfig(hfQuantConfigJson);
-    }
-    const fromConfig = fromConfigQuantizationConfig(configJson);
-    if (fromConfig !== undefined)
-        return fromConfig;
-    return fromDtype(configJson);
-}
-/**
- * The KV-cache's own quantization algorithm — NVIDIA ModelOpt's `kv_cache_quant_algo`, a sibling
- * field to `quant_algo` above, read the same way and from the same object: `hf_quant_config.json`'s
- * `quantization` when the repository carries the file, else `config.json`'s `quantization_config`.
- * Used by the KV-cache memory formula (`compatibility.ts`, task 2.16w round 1) to size a KV cache
- * entry at 1 byte (`FP8`) instead of the engine's own default 2 bytes (`bf16`/`fp16`). `undefined`
- * when neither object carries it — an unquantized KV cache, or a checkpoint this naming rule does
- * not otherwise recognise.
- */
-export function kvCacheQuantAlgo(configJson, hfQuantConfigJson) {
-    if (hfQuantConfigJson !== null) {
-        const quantization = asObject(hfQuantConfigJson.quantization);
-        return quantization === undefined ? undefined : asNonEmptyString(quantization.kv_cache_quant_algo);
-    }
-    const quantizationConfig = asObject(configJson.quantization_config);
-    return quantizationConfig === undefined
-        ? undefined
-        : asNonEmptyString(quantizationConfig.kv_cache_quant_algo);
-}
-/**
- * What the naming rule actually saw when it did not recognise a format, for error messages only —
- * never consulted to decide the format itself (that is `quantizationFormat`'s job alone). Mirrors
- * `quantizationFormat`'s own step order, so the two must be kept in sync.
- */
-export function describeUnrecognizedQuantization(configJson, hfQuantConfigJson) {
-    if (hfQuantConfigJson !== null) {
-        const quantization = asObject(hfQuantConfigJson.quantization);
-        const quantAlgo = quantization?.quant_algo;
-        return `hf_quant_config.json quantization.quant_algo=${JSON.stringify(quantAlgo ?? null)}`;
-    }
-    const quantizationConfig = asObject(configJson.quantization_config);
-    if (quantizationConfig !== undefined) {
-        const quantMethod = quantizationConfig.quant_method;
-        if (quantMethod === 'modelopt') {
-            return `config.json quantization_config.quant_method="modelopt" quant_algo=${JSON.stringify(quantizationConfig.quant_algo ?? null)}`;
-        }
-        if (quantMethod === 'fp8') {
-            return `config.json quantization_config.quant_method="fp8" weight_block_size=${JSON.stringify(quantizationConfig.weight_block_size ?? null)}`;
-        }
-        return `config.json quantization_config.quant_method=${JSON.stringify(quantMethod ?? null)}`;
-    }
-    return `config.json dtype=${JSON.stringify(configDtype(configJson) ?? null)}`;
-}
-/** What NVIDIA ModelOpt calls a checkpoint whose layers carry different formats. */
-export const MIXED_PRECISION = 'mixed_precision';
-/**
- * The formats a checkpoint actually needs the engine to load. One format for every checkpoint but a
- * ModelOpt `MIXED_PRECISION` one, whose `hf_quant_config.json` lists a `quant_algo` per layer in
- * `quantization.quantized_layers` (e.g. `NVFP4` + `FP8`, or `W4A16_NVFP4` + `FP8`): every distinct
- * per-layer format, named by the same rule as a whole-checkpoint `quant_algo`. An empty list means
- * the mixed checkpoint names no per-layer format at all, which the caller must reject.
- */
-export function quantizationComponents(format, hfQuantConfigJson) {
-    if (format !== MIXED_PRECISION)
-        return [format];
-    const quantization = hfQuantConfigJson === null ? undefined : asObject(hfQuantConfigJson.quantization);
-    const layers = quantization === undefined ? undefined : asObject(quantization.quantized_layers);
-    if (layers === undefined)
-        return [];
-    const formats = new Set();
-    for (const layer of Object.values(layers)) {
-        const quantAlgo = asNonEmptyString(asObject(layer)?.quant_algo);
-        if (quantAlgo === undefined)
-            return [];
-        formats.add(renameFp8PbWo(quantAlgo));
-    }
-    return [...formats].sort();
-}
-//# sourceMappingURL=quant-format.js.map

package/dist/runtime/tensorrt-llm/quant-format.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"quant-format.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/quant-format.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAKH,MAAM,SAAS,GAAG,WAAW,CAAA;AAC7B,MAAM,gBAAgB,GAAG,kBAAkB,CAAA;AAE3C,SAAS,QAAQ,CAAC,KAAc;IAC9B,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QACzE,CAAC,CAAE,KAAoB;QACvB,CAAC,CAAC,SAAS,CAAA;AACf,CAAC;AAED,SAAS,gBAAgB,CAAC,KAAc;IACtC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAA;AAC1E,CAAC;AAED,gHAAgH;AAChH,SAAS,aAAa,CAAC,SAAiB;IACtC,MAAM,KAAK,GAAG,SAAS,CAAC,WAAW,EAAE,CAAA;IACrC,OAAO,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,gBAAgB,CAAC,CAAC,CAAC,KAAK,CAAA;AACvD,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,gBAAgB,CAAC,KAAkC;IACjE,OAAO,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAA;AACxE,CAAC;AAED;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,iBAA6B;IACtD,MAAM,YAAY,GAAG,QAAQ,CAAC,iBAAiB,CAAC,YAAY,CAAC,CAAA;IAC7D,MAAM,SAAS,GAAG,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,gBAAgB,CAAC,YAAY,CAAC,UAAU,CAAC,CAAA;IACpG,OAAO,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,aAAa,CAAC,SAAS,CAAC,CAAA;AAClE,CAAC;AAED;;;;;GAKG;AACH,SAAS,4BAA4B,CAAC,UAAsB;IAC1D,MAAM,kBAAkB,GAAG,QAAQ,CAAC,UAAU,CAAC,mBAAmB,CAAC,CAAA;IACnE,IAAI,kBAAkB,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IACtD,MAAM,WAAW,GAAG,kBAAkB,CAAC,YAAY,CAAA;IACnD,IAAI,WAAW,KAAK,UAAU,EAAE,CAAC;QAC/B,MAAM,SAAS,GAAG,gBAAgB,CAAC,kBAAkB,CAAC,UAAU,CAAC,CAAA;QACjE,OAAO,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,aAAa,CAAC,SAAS,CAAC,CAAA;IAClE,CAAC;IACD,IAAI,WAAW,KAAK,KAAK,EAAE,CAAC;QAC1B,MAAM,SAAS,GAAG,kBAAkB,CAAC,iBAAiB,CAAA;QACtD,MAAM,gBAAgB,GACpB,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,IAAI,SAAS,CAAC,CAAC,CAAC,KAAK,GAAG,IAAI,SAAS,CAAC,CAAC,CAAC,KAAK,GAAG,CAAA;QACpG,6EAA6E;QAC7E,gFAAgF;QAChF,OAAO,gBAAgB,CAAC,CAAC,CAAC,gBAAgB,CAAC,CAAC,CAAC,IAAI,CAAA;IACnD,CAAC;IACD,IAAI,WAAW,KAAK,OAAO;QAAE,OAAO,OAAO,CAAA;IAC3C,iEAAiE;IACjE,OAAO,IAAI,CAAA;AACb,CAAC;AAED;;;;GAIG;AACH,SAAS,WAAW,CAAC,UAAsB;IACzC,MAAM,UAAU,GAAG,QAAQ,CAAC,UAAU,CAAC,WAAW,CAAC,CAAA;IACnD,OAAO,CACL,gBAAgB,CAAC,UAAU,CAAC,KAAK,CAAC;QAClC,gBAAgB,CAAC,UAAU,CAAC,WAAW,CAAC;QACxC,CAAC,UAAU,KAAK,SAAS;YACvB,CAAC,CAAC,SAAS;YACX,CAAC,CAAC,CAAC,gBAAgB,CAAC,UAAU,CAAC,KAAK,CAAC,IAAI,gBAAgB,CAAC,UAAU,CAAC,WAAW,CAAC,CAAC,CAAC,CACtF,CAAA;AACH,CAAC;AAED,mDAAmD;AACnD,SAAS,SAAS,CAAC,UAAsB;IACvC,MAAM,KAAK,GAAG,WAAW,CAAC,UAAU,CAAC,CAAA;IACrC,IAAI,KAAK,KAAK,UAAU;QAAE,OAAO,MAAM,CAAA;IACvC,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,MAAM,CAAA;IACtC,OAAO,IAAI,CAAA;AACb,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,kBAAkB,CAChC,UAAsB,EACtB,iBAAoC;IAEpC,IAAI,iBAAiB,KAAK,IAAI,EAAE,CAAC;QAC/B,OAAO,iBAAiB,CAAC,iBAAiB,CAAC,CAAA;IAC7C,CAAC;IACD,MAAM,UAAU,GAAG,4BAA4B,CAAC,UAAU,CAAC,CAAA;IAC3D,IAAI,UAAU,KAAK,SAAS;QAAE,OAAO,UAAU,CAAA;IAC/C,OAAO,SAAS,CAAC,UAAU,CAAC,CAAA;AAC9B,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,gBAAgB,CAC9B,UAAsB,EACtB,iBAAoC;IAEpC,IAAI,iBAAiB,KAAK,IAAI,EAAE,CAAC;QAC/B,MAAM,YAAY,GAAG,QAAQ,CAAC,iBAAiB,CAAC,YAAY,CAAC,CAAA;QAC7D,OAAO,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,gBAAgB,CAAC,YAAY,CAAC,mBAAmB,CAAC,CAAA;IACpG,CAAC;IACD,MAAM,kBAAkB,GAAG,QAAQ,CAAC,UAAU,CAAC,mBAAmB,CAAC,CAAA;IACnE,OAAO,kBAAkB,KAAK,SAAS;QACrC,CAAC,CAAC,SAAS;QACX,CAAC,CAAC,gBAAgB,CAAC,kBAAkB,CAAC,mBAAmB,CAAC,CAAA;AAC9D,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,gCAAgC,CAC9C,UAAsB,EACtB,iBAAoC;IAEpC,IAAI,iBAAiB,KAAK,IAAI,EAAE,CAAC;QAC/B,MAAM,YAAY,GAAG,QAAQ,CAAC,iBAAiB,CAAC,YAAY,CAAC,CAAA;QAC7D,MAAM,SAAS,GAAG,YAAY,EAAE,UAAU,CAAA;QAC1C,OAAO,gDAAgD,IAAI,CAAC,SAAS,CAAC,SAAS,IAAI,IAAI,CAAC,EAAE,CAAA;IAC5F,CAAC;IACD,MAAM,kBAAkB,GAAG,QAAQ,CAAC,UAAU,CAAC,mBAAmB,CAAC,CAAA;IACnE,IAAI,kBAAkB,KAAK,SAAS,EAAE,CAAC;QACrC,MAAM,WAAW,GAAG,kBAAkB,CAAC,YAAY,CAAA;QACnD,IAAI,WAAW,KAAK,UAAU,EAAE,CAAC;YAC/B,OAAO,sEAAsE,IAAI,CAAC,SAAS,CAAC,kBAAkB,CAAC,UAAU,IAAI,IAAI,CAAC,EAAE,CAAA;QACtI,CAAC;QACD,IAAI,WAAW,KAAK,KAAK,EAAE,CAAC;YAC1B,OAAO,wEAAwE,IAAI,CAAC,SAAS,CAAC,kBAAkB,CAAC,iBAAiB,IAAI,IAAI,CAAC,EAAE,CAAA;QAC/I,CAAC;QACD,OAAO,gDAAgD,IAAI,CAAC,SAAS,CAAC,WAAW,IAAI,IAAI,CAAC,EAAE,CAAA;IAC9F,CAAC;IACD,OAAO,qBAAqB,IAAI,CAAC,SAAS,CAAC,WAAW,CAAC,UAAU,CAAC,IAAI,IAAI,CAAC,EAAE,CAAA;AAC/E,CAAC;AAED,oFAAoF;AACpF,MAAM,CAAC,MAAM,eAAe,GAAG,iBAAiB,CAAA;AAEhD;;;;;;GAMG;AACH,MAAM,UAAU,sBAAsB,CAAC,MAAc,EAAE,iBAAoC;IACzF,IAAI,MAAM,KAAK,eAAe;QAAE,OAAO,CAAC,MAAM,CAAC,CAAA;IAC/C,MAAM,YAAY,GAAG,iBAAiB,KAAK,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,QAAQ,CAAC,iBAAiB,CAAC,YAAY,CAAC,CAAA;IACtG,MAAM,MAAM,GAAG,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,QAAQ,CAAC,YAAY,CAAC,gBAAgB,CAAC,CAAA;IAC/F,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,EAAE,CAAA;IACnC,MAAM,OAAO,GAAG,IAAI,GAAG,EAAU,CAAA;IACjC,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;QAC1C,MAAM,SAAS,GAAG,gBAAgB,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,UAAU,CAAC,CAAA;QAC/D,IAAI,SAAS,KAAK,SAAS;YAAE,OAAO,EAAE,CAAA;QACtC,OAAO,CAAC,GAAG,CAAC,aAAa,CAAC,SAAS,CAAC,CAAC,CAAA;IACvC,CAAC;IACD,OAAO,CAAC,GAAG,OAAO,CAAC,CAAC,IAAI,EAAE,CAAA;AAC5B,CAAC"}

package/dist/runtime/tensorrt-llm/registry.d.ts

@@ -1,26 +0,0 @@
-import type { TensorrtLlmModelYmlDocument } from './model-dir.js';
-export interface TensorrtLlmModelEntry {
-    id: string;
-    /** Directory holding `model.yml`. */
-    dir: string;
-    yml: TensorrtLlmModelYmlDocument;
-}
-/** Scan errors are collected rather than thrown: one broken model must not hide the rest of the list. */
-export interface TensorrtLlmScanResult {
-    entries: TensorrtLlmModelEntry[];
-    skipped: Array<{
-        dir: string;
-        error: string;
-    }>;
-}
-export declare class TensorrtLlmModelRegistry {
-    private readonly modelsDir;
-    /**
-     * The models root, or how to learn it at each scan (change `add-tensorrt-llm-windows`, task 2.8): on
-     * Windows it is a folder in Atomic Chat's WSL distribution, known only once that exists — `null`
-     * until then, which lists nothing.
-     */
-    constructor(modelsDir: string | (() => Promise<string | null>));
-    scan(): Promise<TensorrtLlmScanResult>;
-    list(): Promise<TensorrtLlmModelEntry[]>;
-}

package/dist/runtime/tensorrt-llm/registry.js

@@ -1,68 +0,0 @@
-/**
- * The `tensorrt-llm` model registry (task 2.16, spec `tensorrt-llm-models`, "Модель — каталог
- * провайдера с манифестом"): a depth-first scan of `<data>/tensorrt-llm/models`, following the same
- * rules `src/models/registry.ts`'s `ModelRegistry` already uses for llama.cpp so all three surfaces
- * (core, the app, the CLI) see the same ids for the same folders — a directory holding `model.yml`
- * *is* a model and is never descended into, the id is its path relative to the models root with `\`
- * rewritten to `/`, ids sort ascending, and an unreadable or invalid `model.yml` is skipped rather
- * than failing the whole scan (spec "Недокачанный каталог": a directory with files but no
- * `model.yml` is not shown at all).
- *
- * There is no caching here, by design: every call re-walks the directory, so a model the app finishes
- * downloading and writes `model.yml` for appears on the very next `list()` with no restart needed
- * (spec "Модель докачана в app"). `model.yml` itself is a different shape from llama.cpp's
- * (`repository`, `revision`, `files`, `architectures`, `quantization` — no `model_path`), which is
- * why this is its own registry rather than a reuse of the llama.cpp one; `parseTensorrtLlmModelYml`
- * (`model-dir.ts`) is the one parser both this scan and the single-id load-path lookup
- * (`readTensorrtLlmModel`) share, so they can never disagree about what a `model.yml` means.
- */
-import { readdir, readFile } from 'node:fs/promises';
-import { join } from 'node:path';
-import { MODEL_YML, modelIdFromDir } from '../../config/index.js';
-import { parseTensorrtLlmModelYml } from './model-dir.js';
-export class TensorrtLlmModelRegistry {
-    modelsDir;
-    /**
-     * The models root, or how to learn it at each scan (change `add-tensorrt-llm-windows`, task 2.8): on
-     * Windows it is a folder in Atomic Chat's WSL distribution, known only once that exists — `null`
-     * until then, which lists nothing.
-     */
-    constructor(modelsDir) {
-        this.modelsDir = modelsDir;
-    }
-    async scan() {
-        const result = { entries: [], skipped: [] };
-        const root = typeof this.modelsDir === 'string' ? this.modelsDir : await this.modelsDir().catch(() => null);
-        if (root === null)
-            return result;
-        const stack = [root];
-        while (stack.length) {
-            const dir = stack.pop();
-            const ymlPath = join(dir, MODEL_YML);
-            const text = await readFile(ymlPath, 'utf8').catch(() => undefined);
-            if (text !== undefined) {
-                try {
-                    result.entries.push({
-                        id: modelIdFromDir(root, dir),
-                        dir,
-                        yml: parseTensorrtLlmModelYml(text, ymlPath),
-                    });
-                }
-                catch (e) {
-                    result.skipped.push({ dir, error: e.message });
-                }
-                continue; // a model directory is a leaf, whether or not it parsed
-            }
-            const children = await readdir(dir, { withFileTypes: true }).catch(() => []);
-            for (const child of children)
-                if (child.isDirectory())
-                    stack.push(join(dir, child.name));
-        }
-        result.entries.sort((a, b) => (a.id < b.id ? -1 : a.id > b.id ? 1 : 0));
-        return result;
-    }
-    async list() {
-        return (await this.scan()).entries;
-    }
-}
-//# sourceMappingURL=registry.js.map

package/dist/runtime/tensorrt-llm/registry.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"registry.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AACH,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAA;AACpD,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAA;AAChC,OAAO,EAAE,SAAS,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAA;AACjE,OAAO,EAAE,wBAAwB,EAAE,MAAM,gBAAgB,CAAA;AAgBzD,MAAM,OAAO,wBAAwB;IAMN;IAL7B;;;;OAIG;IACH,YAA6B,SAAkD;QAAlD,cAAS,GAAT,SAAS,CAAyC;IAAG,CAAC;IAEnF,KAAK,CAAC,IAAI;QACR,MAAM,MAAM,GAA0B,EAAE,OAAO,EAAE,EAAE,EAAE,OAAO,EAAE,EAAE,EAAE,CAAA;QAClE,MAAM,IAAI,GACR,OAAO,IAAI,CAAC,SAAS,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,SAAS,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAA;QAChG,IAAI,IAAI,KAAK,IAAI;YAAE,OAAO,MAAM,CAAA;QAChC,MAAM,KAAK,GAAa,CAAC,IAAI,CAAC,CAAA;QAC9B,OAAO,KAAK,CAAC,MAAM,EAAE,CAAC;YACpB,MAAM,GAAG,GAAG,KAAK,CAAC,GAAG,EAAY,CAAA;YACjC,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,SAAS,CAAC,CAAA;YACpC,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAA;YACnE,IAAI,IAAI,KAAK,SAAS,EAAE,CAAC;gBACvB,IAAI,CAAC;oBACH,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC;wBAClB,EAAE,EAAE,cAAc,CAAC,IAAI,EAAE,GAAG,CAAC;wBAC7B,GAAG;wBACH,GAAG,EAAE,wBAAwB,CAAC,IAAI,EAAE,OAAO,CAAC;qBAC7C,CAAC,CAAA;gBACJ,CAAC;gBAAC,OAAO,CAAC,EAAE,CAAC;oBACX,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,KAAK,EAAG,CAAW,CAAC,OAAO,EAAE,CAAC,CAAA;gBAC3D,CAAC;gBACD,SAAQ,CAAC,wDAAwD;YACnE,CAAC;YACD,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,GAAG,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAA;YAC5E,KAAK,MAAM,KAAK,IAAI,QAAQ;gBAAE,IAAI,KAAK,CAAC,WAAW,EAAE;oBAAE,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC,CAAA;QAC1F,CAAC;QACD,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAA;QACvE,OAAO,MAAM,CAAA;IACf,CAAC;IAED,KAAK,CAAC,IAAI;QACR,OAAO,CAAC,MAAM,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,OAAO,CAAA;IACpC,CAAC;CACF"}

package/dist/runtime/tensorrt-llm/runtime.d.ts

@@ -1,130 +0,0 @@
-import type { SessionInfo, UnloadResult } from '../../contracts/index.js';
-import type { ModelCapabilities } from '../../models/index.js';
-import type { ManagedLastAttempt, ManagedTextCapabilities, ManagedTextLifecycle } from '../managed-text/index.js';
-import type { CtxIncreaseResult, GpuClaimHook, GpuOccupancy, LocalLoadOptions, LocalRuntime, RecreateResult, SessionRoutePolicy } from '../shared/index.js';
-import type { TensorrtLlmHostFacts } from './host-facts.js';
-import type { ReadyInstallation } from './installation.js';
-import type { TensorrtLlmModel } from './model-dir.js';
-export interface TensorrtLlmRuntimeDeps {
-    /**
-     * The lifecycle over the core's one Docker executor and journal (the startup handle the setup
-     * operation shares), asked at every load: `null` while this host has no docker CLI — a setup that
-     * installs Docker later makes the next load find it, with no restart; rejected, with the cause,
-     * when wiring the executor failed. The same instance every time it resolves non-null.
-     */
-    lifecycle: () => Promise<ManagedTextLifecycle | null>;
-    /** The `ready` installation, its pinned descriptor and image; rejects with why there is none. */
-    readyInstallation: () => Promise<ReadyInstallation>;
-    /** The cards and SELinux, asked right before each load: a card can disappear between two loads. */
-    hostFacts: () => Promise<TensorrtLlmHostFacts>;
-    model: (modelId: string) => Promise<TensorrtLlmModel>;
-    /** The provider's stored settings (`settings.get('tensorrt-llm')`), read at every load. */
-    settings: () => Record<string, unknown>;
-    /**
-     * Core's GPU residency (spec `gpu-residency`), run as the `stopping-previous` stage: frees the chosen
-     * card of every other engine and this provider's other session. Absent, the stage stops only the
-     * other `tensorrt-llm` models, itself.
-     */
-    claimGpu?: GpuClaimHook;
-}
-/**
- * `GET /models/tensorrt-llm/:id/capabilities`: the fields every provider's answer has (all false for
- * this engine — no projector, no embedding, no speculative decoding), and the managed engine's own,
- * which the app gates Agent, tools and attachments on (task 3.8).
- */
-export type TensorrtLlmModelCapabilities = ModelCapabilities & ManagedTextCapabilities & {
-    /** The architecture `model.yml` names, which picked the descriptor's `model_families` entry. */
-    architecture: string | null;
-};
-/** `GET /models/tensorrt-llm/:id/logs`: the loaded container's live tail, or the last failed attempt's. */
-export type TensorrtLlmModelLogs = {
-    model_id: string;
-    source: 'session';
-    generation: string;
-    log_tail: string;
-} | {
-    model_id: string;
-    source: 'last-attempt';
-    generation: string;
-    log_tail: string;
-    error: ManagedLastAttempt['error'];
-    at: number;
-} | {
-    model_id: string;
-    source: null;
-    log_tail: '';
-};
-export declare class TensorrtLlmRuntime implements LocalRuntime {
-    private readonly deps;
-    private current;
-    private closed;
-    /** Removals of the engine in progress: while any holds loads off, every load is refused (M-1). */
-    private loadHolds;
-    /** Deletions of one model in progress (task 2.24): each refuses loads of that model only, counted. */
-    private readonly modelHolds;
-    /** Generations whose GPU claim succeeded: a load only holds its card from then on. */
-    private readonly claimed;
-    /** What each loaded session can do, keyed by model and pinned to the generation it was computed for. */
-    private readonly sessionCapabilities;
-    constructor(deps: TensorrtLlmRuntimeDeps);
-    list(): SessionInfo[];
-    findSession(modelId: string): SessionInfo | undefined;
-    getLoadedModels(): string[];
-    isLoading(modelId: string): boolean;
-    load(modelId: string, opts?: LocalLoadOptions): Promise<SessionInfo>;
-    /**
-     * Throws `MANAGED_STOP_UNCONFIRMED` when Docker will not confirm the stop: the card stays reserved
-     * and the caller is told, rather than handed a success that is not one.
-     */
-    unload(modelId: string): Promise<UnloadResult>;
-    /**
-     * Every model holding a card through this provider — loading, ready, stopping, or stopped without
-     * Docker's confirmation: what a removal of the engine must unload first (task 2.6), through core's
-     * own unload so each model's cross-process claim is released with it (final review M-1).
-     */
-    residentModels(): string[];
-    /**
-     * Refuses every load with `MANAGED_OPERATION_CONFLICT` until the returned release is called (final
-     * review M-1): a removal of the engine holds loads off for as long as it runs, so no load creates
-     * a container on an installation being removed. Holds nest; each release counts once.
-     */
-    holdOffLoads(): () => void;
-    /**
-     * Refuses loads of this one model with `MANAGED_OPERATION_CONFLICT` until the returned release is
-     * called (task 2.24): a deletion holds them off from before it stops the model until its files are
-     * gone, so no load mounts a folder being removed. Other models load as usual. Holds nest.
-     */
-    holdOffModel(modelId: string): () => void;
-    /** Never: the context is fixed when the container starts (design D9, spec "без авто-роста"). */
-    autoIncreaseCtx(_modelId: string): Promise<CtxIncreaseResult>;
-    recreateSession(modelId: string): Promise<RecreateResult>;
-    shutdown(): Promise<void>;
-    routePolicy(modelId: string): SessionRoutePolicy | undefined;
-    /** Answers rather than throws, like every provider's capabilities: all false while nothing resolves. */
-    capabilities(modelId: string): Promise<TensorrtLlmModelCapabilities>;
-    logs(modelId: string): Promise<TensorrtLlmModelLogs>;
-    private assertOpen;
-    private assertNotHeldOff;
-    /**
-     * The card each container holds, for core's residency rule: every model past its claim — loading,
-     * ready, stopping, or stopped without Docker's confirmation, which still holds the card.
-     */
-    gpuOccupancy(): GpuOccupancy[];
-    /**
-     * The `stopping-previous` stage. With core's residency: its claim on `gpuId`, which stops every other
-     * `tensorrt-llm` model (one session at a time) and every other engine on that card, each with a
-     * confirmed exit, and refuses with `GPU_BUSY` while one will not stop. Without it: every other
-     * `tensorrt-llm` model, loading or loaded, stopped with confirmation. Either way the load holds its
-     * card from then on — with residency, from the moment core grants the claim, inside core's turn.
-     */
-    private stopOthers;
-    /**
-     * The lifecycle's `beforeCreate` hook (task 2.16w round 1, finding 1 (Critical)): the pre-launch
-     * check's memory line, re-probed fresh right before the container is actually created — after
-     * `stopPrevious` has run, so a model switch on a single-GPU host sees the card the previous
-     * session just freed, never a snapshot `load()` took before eviction. `MANAGED_PREREQUISITE_
-     * BLOCKED` when the chosen card is no longer on the host at all (the same code a missing card gets
-     * earlier in `load()`); `MODEL_INCOMPATIBLE` when it is still there but does not have room.
-     */
-    private checkMemoryBeforeCreate;
-}