@atomic-chat/core 0.5.1 → 0.13.2

package/dist/runtime/managed-models/compatibility.js

@@ -0,0 +1,594 @@
+/**
+ * The compatibility check every managed engine shares (change `add-vllm-runtime`, design D7; spec
+ * `managed-model-store`, "Проверка совместимости одинакова по форме для всех managed-движков"):
+ * whether a Hugging Face checkpoint can run on one engine on this host, computed entirely from what
+ * the caller already has — `config.json`, `hf_quant_config.json` when the repository carries one, the
+ * revision's file listing, that engine's pinned `RuntimeDescriptor`, the host's `GpuFacts[]` and its
+ * `MemAvailable`/`MemTotal` (`HostMemory`) — before a single byte of the checkpoint is downloaded.
+ *
+ * The skeleton is the same for every engine; two things are the engine's own, passed in as a
+ * `ManagedCheckEngine`: `memoryNeed`, what the checkpoint needs on a card by that engine's own
+ * memory rule (TensorRT-LLM: its KV reserve divided by the free-memory fraction plus its runtime and
+ * activations; vLLM: weights plus KV in bytes plus its overhead), and `checkpointProblems`, shapes of
+ * a supported architecture the engine's loader cannot read. The engine's descriptor decides the rest:
+ * its architectures and its format matrix. Neither the hooks nor the adapter are the same thing — the
+ * check runs with no Docker and no load lifecycle at all.
+ *
+ * This is the pure verdict only: it never touches the network or the filesystem. `selectLaunchGpu` is
+ * exported on its own because a load picks the same card the same way.
+ *
+ * Check order: GGUF is rejected outright before anything else (spec: "for GGUF there is
+ * llama.cpp"), then a curated match's `inventory_digest` is verified — an integrity gate that is
+ * independent of whether the checkpoint would otherwise load, so a tampered curated listing is
+ * reported as `MANAGED_METADATA_INVALID` rather than whatever compatibility error the tampered
+ * files happen to also trigger. Only then: quantization format recognised by the shared naming rule,
+ * architecture supported, the engine's checkpoint problems, format present in this descriptor's own
+ * matrix, the file listing has at least one weight file, the format's compute-capability rule
+ * (minimum and exclusion list) — everything above is genuinely static, true or false regardless of
+ * what else is running on the host — and finally the engine's memory need against the selected
+ * card's free memory.
+ *
+ * `checkCheckpoint` runs every check as one call, for the check route's single live snapshot. The
+ * load path cannot: a pre-launch check runs before the previous session on the very same card is
+ * stopped, so reading that card's free memory then would refuse a same-card model switch outright.
+ * `checkCheckpointFiles` is every check above the memory line, and `checkCheckpointMemory` is the
+ * memory line alone, so a load can run the first before stopping the previous session and the second
+ * after it, against a freshly re-probed card.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+import { inventoryDigest } from '../environment/index.js';
+import { describeUnrecognizedQuantization, isGgufCheckpoint, kvCacheQuantAlgo, MIXED_PRECISION, quantizationComponents, quantizationFormat, } from './quant-format.js';
+/**
+ * `model.layers.12.mlp.experts.7.w1.weight` → `model.layers.*.mlp.experts.*.w1.weight`: a large MoE
+ * index (DeepSeek-V3: ~90k names) folds to a few hundred, and no shape check here looks at indices.
+ */
+export function normalizeWeightName(name) {
+    return name.replace(/(?<=^|\.)\d+(?=\.|$)/g, '*');
+}
+/** Folded and de-duplicated tensor names, in a stable order. */
+export function normalizeWeightNames(names) {
+    return [...new Set([...names].map(normalizeWeightName))].sort();
+}
+/** A safetensors header larger than this is not a header this core reads (real ones are KB to a few MB). */
+const MAX_SAFETENSORS_HEADER_BYTES = 100 * 1024 * 1024;
+/** The header length a `.safetensors` file's first 8 bytes declare (little-endian u64); `undefined` if implausible. */
+export function safetensorsHeaderLength(prefix) {
+    if (prefix.length < 8)
+        return undefined;
+    const length = new DataView(prefix.buffer, prefix.byteOffset, 8).getBigUint64(0, true);
+    return length > 0n && length <= BigInt(MAX_SAFETENSORS_HEADER_BYTES) ? Number(length) : undefined;
+}
+/** The tensor names a parsed safetensors header lists (its keys, minus `__metadata__`); `undefined` if it is not an object. */
+export function safetensorsTensorNames(header) {
+    if (header === null || typeof header !== 'object' || Array.isArray(header))
+        return undefined;
+    return Object.keys(header).filter((key) => key !== '__metadata__');
+}
+/**
+ * The card a load runs on, and the card `POST /models/:provider/check` computes against when the
+ * caller names none — one function for both, so a verdict and a launch can never be about different
+ * cards (spec `tensorrt-llm-runtime` "Выбор карты и настройки провайдера", design D12b). `gpu_id`
+ * when given and present on the host; otherwise the card with the most FREE memory right now, ties
+ * broken by the most TOTAL memory, and a full tie by the order `nvidia-smi` lists the cards in (its
+ * own index, stable on a host), so equal cards always resolve the same way. `null` only when the host
+ * has no GPU at all.
+ *
+ * Free memory, not total: on a desktop with two equal cards the first is usually holding the
+ * desktop and a browser, and "the biggest card" would send the model there to run out of memory
+ * while the second card sits idle. Placement may therefore change between loads as the cards' load
+ * changes; the engine cache is keyed by model, not by card, so that only moves where it runs.
+ * Compute capability is deliberately not a filter here: the spec's rule is memory only, and a card
+ * the checkpoint's format cannot run on gets an honest `MODEL_INCOMPATIBLE` naming the other cards
+ * it would fit on (`fits_other_gpus`).
+ *
+ * A unified-memory card (`total_vram_bytes: null`) ranks by the host's memory — `MemAvailable` as
+ * its free memory and `MemTotal` as its total — the same figure `freeMemoryBytes` compares its
+ * memory need against.
+ */
+export function selectLaunchGpu(gpus, host, gpuId) {
+    if (gpuId !== undefined) {
+        const requested = gpus.find((gpu) => gpu.gpu_id === gpuId);
+        if (requested !== undefined)
+            return requested;
+    }
+    if (gpus.length === 0)
+        return null;
+    // Strictly greater only: on a full tie the earlier card stays, i.e. nvidia-smi's own order.
+    const ranksAbove = (gpu, best) => {
+        const free = freeMemoryBytes(gpu, host) - freeMemoryBytes(best, host);
+        return free !== 0 ? free > 0 : totalMemoryBytes(gpu, host) > totalMemoryBytes(best, host);
+    };
+    return gpus.reduce((best, gpu) => (ranksAbove(gpu, best) ? gpu : best));
+}
+const SAFETENSORS_SUFFIX = '.safetensors';
+const LEGACY_WEIGHT_SUFFIXES = ['.bin', '.pth'];
+/** `model.safetensors`, or a sharded `model-00001-of-00003.safetensors` (any digit width). */
+const MODEL_SHARD_PATTERN = /^model(-\d+-of-\d+)?\.safetensors$/i;
+/** A file at the root of the listing: a variant subfolder, an ONNX export, ... never counts as a weight. */
+function isRootLevel(path) {
+    return !path.includes('/');
+}
+function isPreferredShard(path) {
+    return isRootLevel(path) && MODEL_SHARD_PATTERN.test(path);
+}
+function isAnySafetensors(path) {
+    return isRootLevel(path) && path.toLowerCase().endsWith(SAFETENSORS_SUFFIX);
+}
+function isLegacyWeightFile(path) {
+    if (!isRootLevel(path))
+        return false;
+    const lower = path.toLowerCase();
+    return LEGACY_WEIGHT_SUFFIXES.some((suffix) => lower.endsWith(suffix));
+}
+const sumSizes = (files) => files.reduce((sum, file) => sum + file.size, 0);
+/**
+ * The files that count as checkpoint weights. Root-level files only. Prefers the standard
+ * `model[-NNNNN-of-MMMMM].safetensors` shard naming when present — which also excludes a
+ * `consolidated*.safetensors` sitting next to it (Mistral's own releases ship both, covering the
+ * exact same weights, for their own inference stack; counting both would double the checkpoint's
+ * real size) — because that naming alone identifies the checkpoint's real weights unambiguously.
+ * When no file matches that preferred naming, every other root-level `*.safetensors` file is
+ * selected instead, `consolidated*` included: this is the *only* branch a `consolidated`-only
+ * repository (no HF shard naming at all) ever reaches, so its `consolidated.safetensors` has to
+ * count here. Only when there is no safetensors file at all does a legacy `*.bin`/`*.pth` checkpoint
+ * count, so a checkpoint this engine cannot load (it only reads safetensors) still gets an honest,
+ * non-zero total rather than a silent `0` that would let `checkModelCompatibilityFiles` report `ok`
+ * on any card. No file matches any of these rules for an empty selection, which `weightBytes` turns
+ * into `0` and `checkModelCompatibilityFiles` itself turns into `MODEL_INCOMPATIBLE`, never a false
+ * `ok`.
+ */
+function selectWeightFiles(files) {
+    const preferredShards = files.filter((file) => isPreferredShard(file.path));
+    if (preferredShards.length > 0)
+        return preferredShards;
+    const anySafetensors = files.filter((file) => isAnySafetensors(file.path));
+    if (anySafetensors.length > 0)
+        return anySafetensors;
+    return files.filter((file) => isLegacyWeightFile(file.path));
+}
+/** The weight files whose safetensors headers hold the checkpoint's tensor names (legacy `.bin`/`.pth` excluded). */
+export function weightSafetensorsFiles(files) {
+    return selectWeightFiles(files).filter((file) => file.path.toLowerCase().endsWith(SAFETENSORS_SUFFIX));
+}
+/** Weight bytes for the memory check: the sum of `selectWeightFiles(files)`. */
+export function weightBytes(files) {
+    return sumSizes(selectWeightFiles(files));
+}
+/**
+ * A finite, positive number read from `configJson[key]`; `undefined` for anything else (missing,
+ * the wrong type, zero, negative, `NaN`/`Infinity`) — every field the KV-cache formula below needs
+ * is a strictly positive count or dimension, and treating `0` or a negative value as "present" would
+ * size a KV cache of zero or negative bytes, silently passing a check that should have fallen back.
+ */
+export function positiveNumberField(configJson, key) {
+    const value = configJson[key];
+    return typeof value === 'number' && Number.isFinite(value) && value > 0 ? value : undefined;
+}
+/**
+ * The architecture fields a transformer's KV-cache size is computed from, with the two standard
+ * fallbacks: `num_key_value_heads` falls back to `num_attention_heads` (plain multi-head attention —
+ * every attention head is its own KV head, the GQA/MQA field simply absent) and `head_dim` falls
+ * back to `hidden_size / num_attention_heads` (the standard derivation, for a config that does not
+ * spell `head_dim` out explicitly). `undefined` when `config.json` lacks what even the fallbacks
+ * need — `kvCacheBytes` then has nothing to compute from, and `kvCacheReserveBytes` falls back to
+ * the older weight-proportional rule.
+ */
+export function readKvCacheShape(rootConfigJson) {
+    // A VLM-style config (Qwen3.5+, Gemma 4, Mistral 3) keeps the language model's shape on
+    // `text_config`; the top level then has no `num_hidden_layers` at all.
+    const textConfig = rootConfigJson.text_config;
+    const configJson = positiveNumberField(rootConfigJson, 'num_hidden_layers') === undefined &&
+        typeof textConfig === 'object' &&
+        textConfig !== null &&
+        !Array.isArray(textConfig)
+        ? textConfig
+        : rootConfigJson;
+    const numHiddenLayers = attentionLayerCount(configJson);
+    const numAttentionHeads = positiveNumberField(configJson, 'num_attention_heads');
+    const numKeyValueHeads = positiveNumberField(configJson, 'num_key_value_heads') ?? numAttentionHeads;
+    const hiddenSize = positiveNumberField(configJson, 'hidden_size');
+    const headDim = positiveNumberField(configJson, 'head_dim') ??
+        (hiddenSize !== undefined && numAttentionHeads !== undefined ? hiddenSize / numAttentionHeads : undefined);
+    if (numHiddenLayers === undefined || numKeyValueHeads === undefined || headDim === undefined) {
+        return undefined;
+    }
+    return { numHiddenLayers, numKeyValueHeads, headDim };
+}
+/**
+ * Layers that keep a per-token KV cache. A hybrid config lists `layer_types`; its
+ * `linear_attention` layers (Qwen3.5's Gated DeltaNet) hold a fixed-size state instead, so only the
+ * rest count. Sliding-window layers still count in full, which over-reserves for them on purpose.
+ */
+function attentionLayerCount(configJson) {
+    const total = positiveNumberField(configJson, 'num_hidden_layers');
+    const layerTypes = configJson.layer_types;
+    if (total === undefined || !Array.isArray(layerTypes) || layerTypes.length !== total)
+        return total;
+    const attention = layerTypes.filter((type) => type !== 'linear_attention').length;
+    return attention > 0 ? attention : total;
+}
+/**
+ * The KV cache's own footprint for one sequence at `contextLength` tokens:
+ * `2 × num_hidden_layers × num_key_value_heads × head_dim × kv_dtype_bytes × context_length` — one
+ * K tensor and one V tensor (the leading `2`) per layer, each holding `context_length` tokens of
+ * `num_key_value_heads × head_dim` values at `kv_dtype_bytes` each. `kv_dtype_bytes` is `1` when
+ * `kvCacheQuantAlgo` names `FP8`, else `2` (`bf16`/`fp16`, this engine's other supported weight
+ * dtypes). `undefined` when `config.json` lacks the architecture fields `readKvCacheShape` needs.
+ */
+export function kvCacheBytes(configJson, hfQuantConfigJson, contextLength) {
+    const shape = readKvCacheShape(configJson);
+    if (shape === undefined)
+        return undefined;
+    const kvDtypeBytes = kvCacheQuantAlgo(configJson, hfQuantConfigJson)?.toUpperCase() === 'FP8' ? 1 : 2;
+    return 2 * shape.numHiddenLayers * shape.numKeyValueHeads * shape.headDim * kvDtypeBytes * contextLength;
+}
+/** A card with no VRAM figure of its own (GB10/DGX Spark, design D13): its memory is the host's. */
+export function isUnifiedMemory(gpu) {
+    return gpu.total_vram_bytes === null;
+}
+/** Free memory to compare against for one card: `MemAvailable` for a unified-memory card (design D13). */
+export function freeMemoryBytes(gpu, host) {
+    return isUnifiedMemory(gpu) ? host.availableBytes : (gpu.free_vram_bytes ?? 0);
+}
+/** A card's size, for ranking only: `MemTotal` for a unified-memory card (design D13). */
+export function totalMemoryBytes(gpu, host) {
+    return isUnifiedMemory(gpu) ? host.totalBytes : (gpu.total_vram_bytes ?? 0);
+}
+/**
+ * Numeric compare of two `"major.minor[.patch...]"` compute-capability strings, `a - b`'s sign.
+ * `NaN` when either side has a non-numeric component (an unparseable `compute_capability`) —
+ * callers must never treat a `NaN` result as "not less than", which would fail open.
+ */
+function compareComputeCapability(a, b) {
+    const partsA = a.split('.').map((part) => Number.parseInt(part, 10));
+    const partsB = b.split('.').map((part) => Number.parseInt(part, 10));
+    const length = Math.max(partsA.length, partsB.length);
+    for (let index = 0; index < length; index += 1) {
+        const da = partsA[index] ?? 0;
+        const db = partsB[index] ?? 0;
+        if (da !== db)
+            return da - db;
+    }
+    return 0;
+}
+/**
+ * Whether `actual` meets `min`. `false` whenever `compareComputeCapability` cannot parse either
+ * side (`NaN`): an unparseable compute capability is treated as incompatible, never as "compatible
+ * by default" — `NaN < 0` is `false`, so comparing the raw sign directly would fail open.
+ */
+function computeCapabilityAtLeast(actual, min) {
+    const compared = compareComputeCapability(actual, min);
+    return !Number.isNaN(compared) && compared >= 0;
+}
+/**
+ * The descriptor's rule for `format`. A `mixed_precision` checkpoint needs every one of its per-layer
+ * formats (`quantizationComponents`) to have a row, and a card has to clear all of them: the highest
+ * minimum and every exclusion. Any per-layer format without a row fails the whole checkpoint closed.
+ */
+function formatRequirement(descriptor, format, hfQuantConfigJson) {
+    const components = quantizationComponents(format, hfQuantConfigJson);
+    if (components.length === 0)
+        return { ok: false, missing: null };
+    let minimum;
+    const excluded = new Set();
+    for (const component of components) {
+        const row = descriptor.quantization.find((entry) => entry.format === component);
+        if (row === undefined)
+            return { ok: false, missing: component };
+        if (minimum === undefined || compareComputeCapability(row.min_compute_capability, minimum) > 0) {
+            minimum = row.min_compute_capability;
+        }
+        for (const capability of row.excluded_compute_capabilities)
+            excluded.add(capability);
+    }
+    return {
+        ok: true,
+        requirement: {
+            min_compute_capability: minimum,
+            excluded_compute_capabilities: [...excluded],
+        },
+    };
+}
+/** `mixed_precision (fp8 + nvfp4)` for a mixed checkpoint, the format itself otherwise. */
+function formatLabel(format, hfQuantConfigJson) {
+    if (format !== MIXED_PRECISION)
+        return format;
+    const components = quantizationComponents(format, hfQuantConfigJson);
+    return components.length === 0 ? format : `${format} (${components.join(' + ')})`;
+}
+/** Whether a checkpoint with `requirement` is loadable on `gpu`. */
+function formatAllowedOnGpu(requirement, gpu) {
+    if (!computeCapabilityAtLeast(gpu.compute_capability, requirement.min_compute_capability))
+        return false;
+    return !requirement.excluded_compute_capabilities.includes(gpu.compute_capability);
+}
+/**
+ * Every other GPU (not `selected`) the checkpoint would fit on: architecture-independent, the
+ * format's CC rule plus free memory, each card measured by its own kind of reserve (`needOn`).
+ */
+function fitsOtherGpus(gpus, selected, requirement, needOn, hostMemory) {
+    if (requirement === null)
+        return [];
+    return gpus
+        .filter((gpu) => gpu.gpu_id !== selected.gpu_id)
+        .filter((gpu) => formatAllowedOnGpu(requirement, gpu))
+        .filter((gpu) => needOn(gpu) <= freeMemoryBytes(gpu, hostMemory))
+        .map((gpu) => gpu.gpu_id);
+}
+/** `exactOptionalPropertyTypes` needs `sha256` omitted, not set to `undefined`, when there is none. */
+function toInventoryFile(file) {
+    return file.sha256 === null
+        ? { path: file.path, bytes: file.size }
+        : { path: file.path, bytes: file.size, sha256: file.sha256 };
+}
+function readArchitectures(configJson) {
+    const value = configJson.architectures;
+    return Array.isArray(value) ? value.filter((entry) => typeof entry === 'string') : [];
+}
+function buildCompatibility(context, verdict, fitsOther, kvReserveBasis) {
+    return {
+        architectures: context.architectures,
+        quantization_format: context.quantizationFormat,
+        weight_bytes: context.weightBytesTotal,
+        checked_gpu_id: context.selected.gpu_id,
+        curated: context.curated,
+        unified_memory: isUnifiedMemory(context.selected),
+        fits_other_gpus: fitsOther,
+        ...(kvReserveBasis === undefined ? {} : { kv_reserve_basis: kvReserveBasis }),
+        verdict,
+    };
+}
+/**
+ * Throws `AtomicCoreError` for two caller-input problems that are never a fact about the checkpoint
+ * itself: `INVALID_ARGUMENT` when the file listing names `hf_quant_config.json` while
+ * `hf_quant_config_json` is `null` (the caller said the file exists but did not send its content, so
+ * the format naming rule cannot be trusted), and `MANAGED_PREREQUISITE_BLOCKED` when the host has no
+ * GPU at all (`gpus` empty — the same code, and the same host condition, `runtime.ts`'s own
+ * `selectLaunchGpu` check already answers with for a real load; task 2.16w round 1, finding 11:
+ * an absent or failing `nvidia-smi` is a missing prerequisite, not a malformed request).
+ * `inventoryDigest` can also throw `AtomicCoreError('MANAGED_METADATA_INVALID', …)` for a curated
+ * match whose file listing is itself malformed (an empty or repeated path, a NUL byte, a
+ * non-integer size — see `src/runtime/environment/inventory.ts`), before this function gets a
+ * chance to compare digests. Every other input, however incompatible, is a normal `verdict.ok: false`
+ * answer, never a thrown error (design D12: the point of this check is to hand back numbers, not to
+ * fail the request).
+ *
+ * Every check above the memory line: GGUF, the curated digest, the quantization format, the
+ * architecture, the format's presence in this descriptor and its compute-capability rule. None of
+ * these can change by evicting whatever session currently holds the selected card, so all of them
+ * run before `stopPrevious` on the load path (task 2.16w round 1, finding 1). `fits_other_gpus` on a
+ * compute-capability failure is still computed here (`memory` sizes the reserve needed to report it),
+ * using whichever `gpus`/`hostMemory` snapshot the caller passed in — informational only,
+ * about *other* cards the eviction race does not touch.
+ */
+export function checkCheckpointFiles(input, descriptor, gpus, hostMemory, engine) {
+    if (input.files.some((file) => file.path === 'hf_quant_config.json') &&
+        input.hf_quant_config_json === null) {
+        throw new AtomicCoreError('INVALID_ARGUMENT', 'The file listing includes hf_quant_config.json but hf_quant_config_json was not provided.');
+    }
+    const selected = selectLaunchGpu(gpus, hostMemory, input.gpu_id);
+    if (selected === null) {
+        throw new AtomicCoreError('MANAGED_PREREQUISITE_BLOCKED', `No NVIDIA GPU was found on this machine, so ${engine.engineId} compatibility cannot be checked.`);
+    }
+    const architectures = readArchitectures(input.config_json);
+    const gguf = isGgufCheckpoint(input.files);
+    const format = gguf ? null : quantizationFormat(input.config_json, input.hf_quant_config_json);
+    const weightBytesTotal = weightBytes(input.files);
+    const curatedEntry = descriptor.curated_models.find((model) => model.repository === input.repository && model.revision === input.revision);
+    const context = {
+        architectures,
+        quantizationFormat: format,
+        weightBytesTotal,
+        selected,
+        curated: curatedEntry !== undefined,
+    };
+    // Integrity gate first: a curated repository/revision match whose files do not hash to the
+    // pinned inventory_digest is reported as tampered metadata, not as whatever ordinary
+    // compatibility error those files happen to also trigger.
+    if (curatedEntry !== undefined) {
+        const actualDigest = inventoryDigest(input.files.map(toInventoryFile));
+        if (actualDigest !== curatedEntry.inventory_digest) {
+            return {
+                ok: false,
+                verdict: buildCompatibility({ ...context, curated: false }, {
+                    ok: false,
+                    error: {
+                        code: 'MANAGED_METADATA_INVALID',
+                        message: 'The submitted file listing does not match the curated inventory digest.',
+                        details: `expected=${curatedEntry.inventory_digest} actual=${actualDigest}`,
+                    },
+                }, []),
+            };
+        }
+    }
+    if (gguf) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: `GGUF checkpoints are not supported by ${engine.engineId}; use the llama.cpp provider for GGUF.`,
+                },
+            }, []),
+        };
+    }
+    if (format === null) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: `Unsupported quantization format: ${describeUnrecognizedQuantization(input.config_json, input.hf_quant_config_json)}.`,
+                    details: `repository=${input.repository} revision=${input.revision}`,
+                },
+            }, []),
+        };
+    }
+    const architecture = architectures.length > 0 ? architectures[0] : undefined;
+    if (architecture === undefined || !descriptor.supported_architectures.includes(architecture)) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: architecture === undefined
+                        ? 'config.json does not declare an architecture.'
+                        : `Unsupported architecture: ${architecture}.`,
+                    ...(architecture !== undefined ? { details: architecture } : {}),
+                },
+            }, []),
+        };
+    }
+    const shapeProblem = engine.checkpointProblems(architecture, input.config_json, input.weight_names);
+    if (shapeProblem !== null) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, { ok: false, error: { code: 'MODEL_INCOMPATIBLE', ...shapeProblem } }, []),
+        };
+    }
+    const resolution = formatRequirement(descriptor, format, input.hf_quant_config_json);
+    if (!resolution.ok) {
+        const missing = resolution.missing;
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: missing === null
+                        ? `Unsupported quantization format: "${format}" does not name the format of its layers.`
+                        : missing === format
+                            ? `Unsupported quantization format: ${engine.engineId} does not support "${format}" (it is not part of this engine's descriptor).`
+                            : `Unsupported quantization format: "${format}" has layers in "${missing}", which ${engine.engineId} does not support (it is not part of this engine's descriptor).`,
+                    details: missing ?? format,
+                },
+            }, []),
+        };
+    }
+    const formatSupport = resolution.requirement;
+    const label = formatLabel(format, input.hf_quant_config_json);
+    if (weightBytesTotal === 0) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: 'No checkpoint weight files were found in the file listing.',
+                },
+            }, []),
+        };
+    }
+    // Known from here on: architecture and format are both fine, so any card whose own CC clears
+    // this format and has room is a real alternative — compute it once and reuse it in both CC-failure
+    // branches below, so a caller whose selected card fails on CC still sees a card that would work
+    // (spec: "report which other host cards it would fit").
+    const checkpoint = {
+        weightBytesTotal,
+        configJson: input.config_json,
+        hfQuantConfigJson: input.hf_quant_config_json,
+    };
+    const needOn = (gpu) => engine.memoryNeed(gpu, checkpoint, hostMemory);
+    const basis = needOn(selected).kvReserveBasis;
+    const fitsOther = fitsOtherGpus(gpus, selected, formatSupport, (gpu) => needOn(gpu).neededBytes, hostMemory);
+    if (!computeCapabilityAtLeast(selected.compute_capability, formatSupport.min_compute_capability)) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: `Format ${label} requires compute capability ${formatSupport.min_compute_capability} or newer.`,
+                    details: `required=${formatSupport.min_compute_capability} actual=${selected.compute_capability}`,
+                },
+            }, fitsOther, basis),
+        };
+    }
+    if (formatSupport.excluded_compute_capabilities.includes(selected.compute_capability)) {
+        return {
+            ok: false,
+            verdict: buildCompatibility(context, {
+                ok: false,
+                error: {
+                    code: 'MODEL_INCOMPATIBLE',
+                    message: `Format ${label} is not supported by this engine release on compute capability ${selected.compute_capability}.`,
+                    details: `format=${format} compute_capability=${selected.compute_capability}`,
+                },
+            }, fitsOther, basis),
+        };
+    }
+    return {
+        ok: true,
+        resolved: {
+            architectures,
+            quantizationFormat: format,
+            weightBytesTotal,
+            selected,
+            curated: context.curated,
+            configJson: input.config_json,
+            hfQuantConfigJson: input.hf_quant_config_json,
+        },
+    };
+}
+/**
+ * The memory line alone (task 2.16w round 1, finding 1): whether `resolved.weightBytesTotal` plus
+ * the KV-cache reserve fits the selected card's free memory. Callers pass a fresh `gpus`/
+ * `hostMemory` snapshot — the load path's `beforeCreate` hook re-probes the host after
+ * `stopPrevious` has run, so this sees whatever that freed, never a snapshot taken before it — and
+ * this function re-reads the free-memory figure for `resolved.selected.gpu_id` from *that* fresh
+ * `gpus`, rather than trusting `resolved.selected`'s own (possibly stale) copy: that stale copy is
+ * only a fallback for the case `gpus` no longer lists the card at all (it disappeared), which
+ * `runtime.ts`'s own `beforeCreate` hook checks for and reports as `MANAGED_PREREQUISITE_BLOCKED`
+ * before ever calling this. This never picks a *different* card — the caller already committed to
+ * this exact one (`checkModelCompatibilityFiles`, or a load's own `selectLaunchGpu`).
+ */
+export function checkCheckpointMemory(resolved, descriptor, gpus, hostMemory, engine) {
+    const selected = gpus.find((gpu) => gpu.gpu_id === resolved.selected.gpu_id) ?? resolved.selected;
+    const context = {
+        architectures: resolved.architectures,
+        quantizationFormat: resolved.quantizationFormat,
+        weightBytesTotal: resolved.weightBytesTotal,
+        selected,
+        curated: resolved.curated,
+    };
+    const checkpoint = {
+        weightBytesTotal: resolved.weightBytesTotal,
+        configJson: resolved.configJson,
+        hfQuantConfigJson: resolved.hfQuantConfigJson,
+    };
+    const needOn = (gpu) => engine.memoryNeed(gpu, checkpoint, hostMemory);
+    const need = needOn(selected);
+    const basis = need.kvReserveBasis;
+    const freeBytes = freeMemoryBytes(selected, hostMemory);
+    const resolution = formatRequirement(descriptor, resolved.quantizationFormat, resolved.hfQuantConfigJson);
+    const fitsOther = fitsOtherGpus(gpus, selected, resolution.ok ? resolution.requirement : null, (gpu) => needOn(gpu).neededBytes, hostMemory);
+    if (need.neededBytes > freeBytes) {
+        return buildCompatibility(context, {
+            ok: false,
+            error: {
+                code: 'MODEL_INCOMPATIBLE',
+                message: 'The checkpoint, the engine runtime memory and the KV-cache reserve do not fit the selected GPU: ' +
+                    `they need ${gib(need.neededBytes)} GiB, ${gib(freeBytes)} GiB is free.`,
+                details: `${need.details} free_bytes=${freeBytes}`,
+            },
+        }, fitsOther, basis);
+    }
+    return buildCompatibility(context, { ok: true }, fitsOther, basis);
+}
+/** Bytes as GiB with one decimal, for a message a person reads. */
+const gib = (bytes) => (bytes / 1024 ** 3).toFixed(1);
+/**
+ * The full verdict: `checkCheckpointFiles` then, if it passed, `checkCheckpointMemory` — for the check
+ * route's single live snapshot, where there is no previous session on the selected card to evict
+ * first. The load path runs the two separately instead (see the file banner).
+ */
+export function checkCheckpoint(input, descriptor, gpus, hostMemory, engine) {
+    const files = checkCheckpointFiles(input, descriptor, gpus, hostMemory, engine);
+    if (!files.ok)
+        return files.verdict;
+    return checkCheckpointMemory(files.resolved, descriptor, gpus, hostMemory, engine);
+}
+//# sourceMappingURL=compatibility.js.map

package/dist/runtime/managed-models/compatibility.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compatibility.js","sourceRoot":"","sources":["../../../src/runtime/managed-models/compatibility.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AAEH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAE1D,OAAO,EAAE,eAAe,EAAE,MAAM,yBAAyB,CAAA;AACzD,OAAO,EACL,gCAAgC,EAChC,gBAAgB,EAChB,gBAAgB,EAChB,eAAe,EACf,sBAAsB,EACtB,kBAAkB,GACnB,MAAM,mBAAmB,CAAA;AAgC1B;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAAC,IAAY;IAC9C,OAAO,IAAI,CAAC,OAAO,CAAC,uBAAuB,EAAE,GAAG,CAAC,CAAA;AACnD,CAAC;AAED,gEAAgE;AAChE,MAAM,UAAU,oBAAoB,CAAC,KAAuB;IAC1D,OAAO,CAAC,GAAG,IAAI,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,mBAAmB,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAA;AACjE,CAAC;AAED,4GAA4G;AAC5G,MAAM,4BAA4B,GAAG,GAAG,GAAG,IAAI,GAAG,IAAI,CAAA;AAEtD,uHAAuH;AACvH,MAAM,UAAU,uBAAuB,CAAC,MAAkB;IACxD,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IACvC,MAAM,MAAM,GAAG,IAAI,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,EAAE,IAAI,CAAC,CAAA;IACtF,OAAO,MAAM,GAAG,EAAE,IAAI,MAAM,IAAI,MAAM,CAAC,4BAA4B,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAA;AACnG,CAAC;AAED,+HAA+H;AAC/H,MAAM,UAAU,sBAAsB,CAAC,MAAe;IACpD,IAAI,MAAM,KAAK,IAAI,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC;QAAE,OAAO,SAAS,CAAA;IAC5F,OAAO,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,KAAK,cAAc,CAAC,CAAA;AACpE,CAAC;AAoBD;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,MAAM,UAAU,eAAe,CAC7B,IAAyB,EACzB,IAAgB,EAChB,KAAc;IAEd,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACxB,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,KAAK,KAAK,CAAC,CAAA;QAC1D,IAAI,SAAS,KAAK,SAAS;YAAE,OAAO,SAAS,CAAA;IAC/C,CAAC;IACD,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAA;IAClC,4FAA4F;IAC5F,MAAM,UAAU,GAAG,CAAC,GAAa,EAAE,IAAc,EAAW,EAAE;QAC5D,MAAM,IAAI,GAAG,eAAe,CAAC,GAAG,EAAE,IAAI,CAAC,GAAG,eAAe,CAAC,IAAI,EAAE,IAAI,CAAC,CAAA;QACrE,OAAO,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,gBAAgB,CAAC,GAAG,EAAE,IAAI,CAAC,GAAG,gBAAgB,CAAC,IAAI,EAAE,IAAI,CAAC,CAAA;IAC3F,CAAC,CAAA;IACD,OAAO,IAAI,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,GAAG,EAAE,EAAE,CAAC,CAAC,UAAU,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAA;AACzE,CAAC;AAED,MAAM,kBAAkB,GAAG,cAAc,CAAA;AACzC,MAAM,sBAAsB,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;AAC/C,8FAA8F;AAC9F,MAAM,mBAAmB,GAAG,qCAAqC,CAAA;AAEjE,4GAA4G;AAC5G,SAAS,WAAW,CAAC,IAAY;IAC/B,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAA;AAC5B,CAAC;AAED,SAAS,gBAAgB,CAAC,IAAY;IACpC,OAAO,WAAW,CAAC,IAAI,CAAC,IAAI,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;AAC5D,CAAC;AAED,SAAS,gBAAgB,CAAC,IAAY;IACpC,OAAO,WAAW,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,kBAAkB,CAAC,CAAA;AAC7E,CAAC;AAED,SAAS,kBAAkB,CAAC,IAAY;IACtC,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC;QAAE,OAAO,KAAK,CAAA;IACpC,MAAM,KAAK,GAAG,IAAI,CAAC,WAAW,EAAE,CAAA;IAChC,OAAO,sBAAsB,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAA;AACxE,CAAC;AAED,MAAM,QAAQ,GAAG,CAAC,KAAgC,EAAU,EAAE,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,IAAI,EAAE,EAAE,CAAC,GAAG,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAA;AAE9G;;;;;;;;;;;;;;;GAeG;AACH,SAAS,iBAAiB,CAAC,KAAgC;IACzD,MAAM,eAAe,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAA;IAC3E,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,eAAe,CAAA;IAEtD,MAAM,cAAc,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAA;IAC1E,IAAI,cAAc,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,cAAc,CAAA;IAEpD,OAAO,KAAK,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAA;AAC9D,CAAC;AAED,qHAAqH;AACrH,MAAM,UAAU,sBAAsB,CAAC,KAAgC;IACrE,OAAO,iBAAiB,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,kBAAkB,CAAC,CAAC,CAAA;AACxG,CAAC;AAED,gFAAgF;AAChF,MAAM,UAAU,WAAW,CAAC,KAAgC;IAC1D,OAAO,QAAQ,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAA;AAC3C,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,UAAsB,EAAE,GAAW;IACrE,MAAM,KAAK,GAAG,UAAU,CAAC,GAAG,CAAC,CAAA;IAC7B,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAA;AAC7F,CAAC;AAQD;;;;;;;;GAQG;AACH,MAAM,UAAU,gBAAgB,CAAC,cAA0B;IACzD,wFAAwF;IACxF,uEAAuE;IACvE,MAAM,UAAU,GAAG,cAAc,CAAC,WAAW,CAAA;IAC7C,MAAM,UAAU,GACd,mBAAmB,CAAC,cAAc,EAAE,mBAAmB,CAAC,KAAK,SAAS;QACtE,OAAO,UAAU,KAAK,QAAQ;QAC9B,UAAU,KAAK,IAAI;QACnB,CAAC,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC;QACxB,CAAC,CAAE,UAAyB;QAC5B,CAAC,CAAC,cAAc,CAAA;IACpB,MAAM,eAAe,GAAG,mBAAmB,CAAC,UAAU,CAAC,CAAA;IACvD,MAAM,iBAAiB,GAAG,mBAAmB,CAAC,UAAU,EAAE,qBAAqB,CAAC,CAAA;IAChF,MAAM,gBAAgB,GAAG,mBAAmB,CAAC,UAAU,EAAE,qBAAqB,CAAC,IAAI,iBAAiB,CAAA;IACpG,MAAM,UAAU,GAAG,mBAAmB,CAAC,UAAU,EAAE,aAAa,CAAC,CAAA;IACjE,MAAM,OAAO,GACX,mBAAmB,CAAC,UAAU,EAAE,UAAU,CAAC;QAC3C,CAAC,UAAU,KAAK,SAAS,IAAI,iBAAiB,KAAK,SAAS,CAAC,CAAC,CAAC,UAAU,GAAG,iBAAiB,CAAC,CAAC,CAAC,SAAS,CAAC,CAAA;IAC5G,IAAI,eAAe,KAAK,SAAS,IAAI,gBAAgB,KAAK,SAAS,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;QAC7F,OAAO,SAAS,CAAA;IAClB,CAAC;IACD,OAAO,EAAE,eAAe,EAAE,gBAAgB,EAAE,OAAO,EAAE,CAAA;AACvD,CAAC;AAED;;;;GAIG;AACH,SAAS,mBAAmB,CAAC,UAAsB;IACjD,MAAM,KAAK,GAAG,mBAAmB,CAAC,UAAU,EAAE,mBAAmB,CAAC,CAAA;IAClE,MAAM,UAAU,GAAG,UAAU,CAAC,WAAW,CAAA;IACzC,IAAI,KAAK,KAAK,SAAS,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,MAAM,KAAK,KAAK;QAAE,OAAO,KAAK,CAAA;IAClG,MAAM,SAAS,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,KAAK,kBAAkB,CAAC,CAAC,MAAM,CAAA;IACjF,OAAO,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,CAAA;AAC1C,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,YAAY,CAC1B,UAAsB,EACtB,iBAAoC,EACpC,aAAqB;IAErB,MAAM,KAAK,GAAG,gBAAgB,CAAC,UAAU,CAAC,CAAA;IAC1C,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IACzC,MAAM,YAAY,GAAG,gBAAgB,CAAC,UAAU,EAAE,iBAAiB,CAAC,EAAE,WAAW,EAAE,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAA;IACrG,OAAO,CAAC,GAAG,KAAK,CAAC,eAAe,GAAG,KAAK,CAAC,gBAAgB,GAAG,KAAK,CAAC,OAAO,GAAG,YAAY,GAAG,aAAa,CAAA;AAC1G,CAAC;AAED,oGAAoG;AACpG,MAAM,UAAU,eAAe,CAAC,GAAa;IAC3C,OAAO,GAAG,CAAC,gBAAgB,KAAK,IAAI,CAAA;AACtC,CAAC;AAED,0GAA0G;AAC1G,MAAM,UAAU,eAAe,CAAC,GAAa,EAAE,IAAgB;IAC7D,OAAO,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,eAAe,IAAI,CAAC,CAAC,CAAA;AAChF,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,gBAAgB,CAAC,GAAa,EAAE,IAAgB;IAC9D,OAAO,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,gBAAgB,IAAI,CAAC,CAAC,CAAA;AAC7E,CAAC;AA0CD;;;;GAIG;AACH,SAAS,wBAAwB,CAAC,CAAS,EAAE,CAAS;IACpD,MAAM,MAAM,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,CAAA;IACpE,MAAM,MAAM,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,CAAA;IACpE,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,CAAC,CAAA;IACrD,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,MAAM,EAAE,KAAK,IAAI,CAAC,EAAE,CAAC;QAC/C,MAAM,EAAE,GAAG,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAA;QAC7B,MAAM,EAAE,GAAG,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAA;QAC7B,IAAI,EAAE,KAAK,EAAE;YAAE,OAAO,EAAE,GAAG,EAAE,CAAA;IAC/B,CAAC;IACD,OAAO,CAAC,CAAA;AACV,CAAC;AAED;;;;GAIG;AACH,SAAS,wBAAwB,CAAC,MAAc,EAAE,GAAW;IAC3D,MAAM,QAAQ,GAAG,wBAAwB,CAAC,MAAM,EAAE,GAAG,CAAC,CAAA;IACtD,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,QAAQ,IAAI,CAAC,CAAA;AACjD,CAAC;AAaD;;;;GAIG;AACH,SAAS,iBAAiB,CACxB,UAA6B,EAC7B,MAAc,EACd,iBAAoC;IAEpC,MAAM,UAAU,GAAG,sBAAsB,CAAC,MAAM,EAAE,iBAAiB,CAAC,CAAA;IACpE,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,CAAA;IAChE,IAAI,OAA2B,CAAA;IAC/B,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAU,CAAA;IAClC,KAAK,MAAM,SAAS,IAAI,UAAU,EAAE,CAAC;QACnC,MAAM,GAAG,GAAG,UAAU,CAAC,YAAY,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAA;QAC/E,IAAI,GAAG,KAAK,SAAS;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,SAAS,EAAE,CAAA;QAC/D,IAAI,OAAO,KAAK,SAAS,IAAI,wBAAwB,CAAC,GAAG,CAAC,sBAAsB,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,CAAC;YAC/F,OAAO,GAAG,GAAG,CAAC,sBAAsB,CAAA;QACtC,CAAC;QACD,KAAK,MAAM,UAAU,IAAI,GAAG,CAAC,6BAA6B;YAAE,QAAQ,CAAC,GAAG,CAAC,UAAU,CAAC,CAAA;IACtF,CAAC;IACD,OAAO;QACL,EAAE,EAAE,IAAI;QACR,WAAW,EAAE;YACX,sBAAsB,EAAE,OAAiB;YACzC,6BAA6B,EAAE,CAAC,GAAG,QAAQ,CAAC;SAC7C;KACF,CAAA;AACH,CAAC;AAED,2FAA2F;AAC3F,SAAS,WAAW,CAAC,MAAc,EAAE,iBAAoC;IACvE,IAAI,MAAM,KAAK,eAAe;QAAE,OAAO,MAAM,CAAA;IAC7C,MAAM,UAAU,GAAG,sBAAsB,CAAC,MAAM,EAAE,iBAAiB,CAAC,CAAA;IACpE,OAAO,UAAU,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,KAAK,UAAU,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAA;AACnF,CAAC;AAED,oEAAoE;AACpE,SAAS,kBAAkB,CAAC,WAA8B,EAAE,GAAa;IACvE,IAAI,CAAC,wBAAwB,CAAC,GAAG,CAAC,kBAAkB,EAAE,WAAW,CAAC,sBAAsB,CAAC;QAAE,OAAO,KAAK,CAAA;IACvG,OAAO,CAAC,WAAW,CAAC,6BAA6B,CAAC,QAAQ,CAAC,GAAG,CAAC,kBAAkB,CAAC,CAAA;AACpF,CAAC;AAED;;;GAGG;AACH,SAAS,aAAa,CACpB,IAAyB,EACzB,QAAkB,EAClB,WAAqC,EACrC,MAAiC,EACjC,UAAsB;IAEtB,IAAI,WAAW,KAAK,IAAI;QAAE,OAAO,EAAE,CAAA;IACnC,OAAO,IAAI;SACR,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,MAAM,CAAC;SAC/C,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,kBAAkB,CAAC,WAAW,EAAE,GAAG,CAAC,CAAC;SACrD,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,MAAM,CAAC,GAAG,CAAC,IAAI,eAAe,CAAC,GAAG,EAAE,UAAU,CAAC,CAAC;SAChE,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,CAAC,CAAA;AAC7B,CAAC;AAED,uGAAuG;AACvG,SAAS,eAAe,CAAC,IAAoB;IAC3C,OAAO,IAAI,CAAC,MAAM,KAAK,IAAI;QACzB,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,IAAI,EAAE;QACvC,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,CAAA;AAChE,CAAC;AAED,SAAS,iBAAiB,CAAC,UAAsB;IAC/C,MAAM,KAAK,GAAG,UAAU,CAAC,aAAa,CAAA;IACtC,OAAO,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,KAAK,EAAmB,EAAE,CAAC,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,CAAA;AACxG,CAAC;AAUD,SAAS,kBAAkB,CACzB,OAAuB,EACvB,OAAsC,EACtC,SAAmB,EACnB,cAA2C;IAE3C,OAAO;QACL,aAAa,EAAE,OAAO,CAAC,aAAa;QACpC,mBAAmB,EAAE,OAAO,CAAC,kBAAkB;QAC/C,YAAY,EAAE,OAAO,CAAC,gBAAgB;QACtC,cAAc,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM;QACvC,OAAO,EAAE,OAAO,CAAC,OAAO;QACxB,cAAc,EAAE,eAAe,CAAC,OAAO,CAAC,QAAQ,CAAC;QACjD,eAAe,EAAE,SAAS;QAC1B,GAAG,CAAC,cAAc,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,cAAc,EAAE,CAAC;QAC7E,OAAO;KACR,CAAA;AACH,CAAC;AAwBD;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,MAAM,UAAU,oBAAoB,CAClC,KAAsB,EACtB,UAA6B,EAC7B,IAAyB,EACzB,UAAsB,EACtB,MAA0B;IAE1B,IACE,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,sBAAsB,CAAC;QAChE,KAAK,CAAC,oBAAoB,KAAK,IAAI,EACnC,CAAC;QACD,MAAM,IAAI,eAAe,CACvB,kBAAkB,EAClB,2FAA2F,CAC5F,CAAA;IACH,CAAC;IAED,MAAM,QAAQ,GAAG,eAAe,CAAC,IAAI,EAAE,UAAU,EAAE,KAAK,CAAC,MAAM,CAAC,CAAA;IAChE,IAAI,QAAQ,KAAK,IAAI,EAAE,CAAC;QACtB,MAAM,IAAI,eAAe,CACvB,8BAA8B,EAC9B,+CAA+C,MAAM,CAAC,QAAQ,mCAAmC,CAClG,CAAA;IACH,CAAC;IAED,MAAM,aAAa,GAAG,iBAAiB,CAAC,KAAK,CAAC,WAAW,CAAC,CAAA;IAC1D,MAAM,IAAI,GAAG,gBAAgB,CAAC,KAAK,CAAC,KAAK,CAAC,CAAA;IAC1C,MAAM,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,kBAAkB,CAAC,KAAK,CAAC,WAAW,EAAE,KAAK,CAAC,oBAAoB,CAAC,CAAA;IAC9F,MAAM,gBAAgB,GAAG,WAAW,CAAC,KAAK,CAAC,KAAK,CAAC,CAAA;IAEjD,MAAM,YAAY,GAAG,UAAU,CAAC,cAAc,CAAC,IAAI,CACjD,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,UAAU,KAAK,KAAK,CAAC,UAAU,IAAI,KAAK,CAAC,QAAQ,KAAK,KAAK,CAAC,QAAQ,CACtF,CAAA;IAED,MAAM,OAAO,GAAmB;QAC9B,aAAa;QACb,kBAAkB,EAAE,MAAM;QAC1B,gBAAgB;QAChB,QAAQ;QACR,OAAO,EAAE,YAAY,KAAK,SAAS;KACpC,CAAA;IAED,2FAA2F;IAC3F,qFAAqF;IACrF,0DAA0D;IAC1D,IAAI,YAAY,KAAK,SAAS,EAAE,CAAC;QAC/B,MAAM,YAAY,GAAG,eAAe,CAAC,KAAK,CAAC,KAAK,CAAC,GAAG,CAAC,eAAe,CAAC,CAAC,CAAA;QACtE,IAAI,YAAY,KAAK,YAAY,CAAC,gBAAgB,EAAE,CAAC;YACnD,OAAO;gBACL,EAAE,EAAE,KAAK;gBACT,OAAO,EAAE,kBAAkB,CACzB,EAAE,GAAG,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,EAC9B;oBACE,EAAE,EAAE,KAAK;oBACT,KAAK,EAAE;wBACL,IAAI,EAAE,0BAA0B;wBAChC,OAAO,EAAE,yEAAyE;wBAClF,OAAO,EAAE,YAAY,YAAY,CAAC,gBAAgB,WAAW,YAAY,EAAE;qBAC5E;iBACF,EACD,EAAE,CACH;aACF,CAAA;QACH,CAAC;IACH,CAAC;IAED,IAAI,IAAI,EAAE,CAAC;QACT,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EAAE,yCAAyC,MAAM,CAAC,QAAQ,wCAAwC;iBAC1G;aACF,EACD,EAAE,CACH;SACF,CAAA;IACH,CAAC;IAED,IAAI,MAAM,KAAK,IAAI,EAAE,CAAC;QACpB,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EAAE,oCAAoC,gCAAgC,CAAC,KAAK,CAAC,WAAW,EAAE,KAAK,CAAC,oBAAoB,CAAC,GAAG;oBAC/H,OAAO,EAAE,cAAc,KAAK,CAAC,UAAU,aAAa,KAAK,CAAC,QAAQ,EAAE;iBACrE;aACF,EACD,EAAE,CACH;SACF,CAAA;IACH,CAAC;IAED,MAAM,YAAY,GAAG,aAAa,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAA;IAC5E,IAAI,YAAY,KAAK,SAAS,IAAI,CAAC,UAAU,CAAC,uBAAuB,CAAC,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC;QAC7F,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EACL,YAAY,KAAK,SAAS;wBACxB,CAAC,CAAC,+CAA+C;wBACjD,CAAC,CAAC,6BAA6B,YAAY,GAAG;oBAClD,GAAG,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;iBACjE;aACF,EACD,EAAE,CACH;SACF,CAAA;IACH,CAAC;IAED,MAAM,YAAY,GAAG,MAAM,CAAC,kBAAkB,CAAC,YAAY,EAAE,KAAK,CAAC,WAAW,EAAE,KAAK,CAAC,YAAY,CAAC,CAAA;IACnG,IAAI,YAAY,KAAK,IAAI,EAAE,CAAC;QAC1B,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE,GAAG,YAAY,EAAE,EAAE,EACrE,EAAE,CACH;SACF,CAAA;IACH,CAAC;IAED,MAAM,UAAU,GAAG,iBAAiB,CAAC,UAAU,EAAE,MAAM,EAAE,KAAK,CAAC,oBAAoB,CAAC,CAAA;IACpF,IAAI,CAAC,UAAU,CAAC,EAAE,EAAE,CAAC;QACnB,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,CAAA;QAClC,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EACL,OAAO,KAAK,IAAI;wBACd,CAAC,CAAC,qCAAqC,MAAM,2CAA2C;wBACxF,CAAC,CAAC,OAAO,KAAK,MAAM;4BAClB,CAAC,CAAC,oCAAoC,MAAM,CAAC,QAAQ,sBAAsB,MAAM,iDAAiD;4BAClI,CAAC,CAAC,qCAAqC,MAAM,oBAAoB,OAAO,YAAY,MAAM,CAAC,QAAQ,iEAAiE;oBAC1K,OAAO,EAAE,OAAO,IAAI,MAAM;iBAC3B;aACF,EACD,EAAE,CACH;SACF,CAAA;IACH,CAAC;IACD,MAAM,aAAa,GAAG,UAAU,CAAC,WAAW,CAAA;IAC5C,MAAM,KAAK,GAAG,WAAW,CAAC,MAAM,EAAE,KAAK,CAAC,oBAAoB,CAAC,CAAA;IAE7D,IAAI,gBAAgB,KAAK,CAAC,EAAE,CAAC;QAC3B,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EAAE,4DAA4D;iBACtE;aACF,EACD,EAAE,CACH;SACF,CAAA;IACH,CAAC;IAED,6FAA6F;IAC7F,mGAAmG;IACnG,gGAAgG;IAChG,wDAAwD;IACxD,MAAM,UAAU,GAA2B;QACzC,gBAAgB;QAChB,UAAU,EAAE,KAAK,CAAC,WAAW;QAC7B,iBAAiB,EAAE,KAAK,CAAC,oBAAoB;KAC9C,CAAA;IACD,MAAM,MAAM,GAAG,CAAC,GAAa,EAAc,EAAE,CAAC,MAAM,CAAC,UAAU,CAAC,GAAG,EAAE,UAAU,EAAE,UAAU,CAAC,CAAA;IAC5F,MAAM,KAAK,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,cAAc,CAAA;IAC7C,MAAM,SAAS,GAAG,aAAa,CAAC,IAAI,EAAE,QAAQ,EAAE,aAAa,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,UAAU,CAAC,CAAA;IAE5G,IAAI,CAAC,wBAAwB,CAAC,QAAQ,CAAC,kBAAkB,EAAE,aAAa,CAAC,sBAAsB,CAAC,EAAE,CAAC;QACjG,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EAAE,UAAU,KAAK,gCAAgC,aAAa,CAAC,sBAAsB,YAAY;oBACxG,OAAO,EAAE,YAAY,aAAa,CAAC,sBAAsB,WAAW,QAAQ,CAAC,kBAAkB,EAAE;iBAClG;aACF,EACD,SAAS,EACT,KAAK,CACN;SACF,CAAA;IACH,CAAC;IAED,IAAI,aAAa,CAAC,6BAA6B,CAAC,QAAQ,CAAC,QAAQ,CAAC,kBAAkB,CAAC,EAAE,CAAC;QACtF,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,kBAAkB,CACzB,OAAO,EACP;gBACE,EAAE,EAAE,KAAK;gBACT,KAAK,EAAE;oBACL,IAAI,EAAE,oBAAoB;oBAC1B,OAAO,EAAE,UAAU,KAAK,kEAAkE,QAAQ,CAAC,kBAAkB,GAAG;oBACxH,OAAO,EAAE,UAAU,MAAM,uBAAuB,QAAQ,CAAC,kBAAkB,EAAE;iBAC9E;aACF,EACD,SAAS,EACT,KAAK,CACN;SACF,CAAA;IACH,CAAC;IAED,OAAO;QACL,EAAE,EAAE,IAAI;QACR,QAAQ,EAAE;YACR,aAAa;YACb,kBAAkB,EAAE,MAAM;YAC1B,gBAAgB;YAChB,QAAQ;YACR,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,UAAU,EAAE,KAAK,CAAC,WAAW;YAC7B,iBAAiB,EAAE,KAAK,CAAC,oBAAoB;SAC9C;KACF,CAAA;AACH,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,qBAAqB,CACnC,QAA4B,EAC5B,UAA6B,EAC7B,IAAyB,EACzB,UAAsB,EACtB,MAA0B;IAE1B,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,QAAQ,CAAC,QAAQ,CAAA;IACjG,MAAM,OAAO,GAAmB;QAC9B,aAAa,EAAE,QAAQ,CAAC,aAAa;QACrC,kBAAkB,EAAE,QAAQ,CAAC,kBAAkB;QAC/C,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB;QAC3C,QAAQ;QACR,OAAO,EAAE,QAAQ,CAAC,OAAO;KAC1B,CAAA;IACD,MAAM,UAAU,GAA2B;QACzC,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB;QAC3C,UAAU,EAAE,QAAQ,CAAC,UAAU;QAC/B,iBAAiB,EAAE,QAAQ,CAAC,iBAAiB;KAC9C,CAAA;IACD,MAAM,MAAM,GAAG,CAAC,GAAa,EAAc,EAAE,CAAC,MAAM,CAAC,UAAU,CAAC,GAAG,EAAE,UAAU,EAAE,UAAU,CAAC,CAAA;IAC5F,MAAM,IAAI,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAA;IAC7B,MAAM,KAAK,GAAG,IAAI,CAAC,cAAc,CAAA;IACjC,MAAM,SAAS,GAAG,eAAe,CAAC,QAAQ,EAAE,UAAU,CAAC,CAAA;IACvD,MAAM,UAAU,GAAG,iBAAiB,CAAC,UAAU,EAAE,QAAQ,CAAC,kBAAkB,EAAE,QAAQ,CAAC,iBAAiB,CAAC,CAAA;IACzG,MAAM,SAAS,GAAG,aAAa,CAC7B,IAAI,EACJ,QAAQ,EACR,UAAU,CAAC,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,WAAW,CAAC,CAAC,CAAC,IAAI,EAC7C,CAAC,GAAG,EAAE,EAAE,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,WAAW,EAChC,UAAU,CACX,CAAA;IAED,IAAI,IAAI,CAAC,WAAW,GAAG,SAAS,EAAE,CAAC;QACjC,OAAO,kBAAkB,CACvB,OAAO,EACP;YACE,EAAE,EAAE,KAAK;YACT,KAAK,EAAE;gBACL,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EACL,kGAAkG;oBAClG,aAAa,GAAG,CAAC,IAAI,CAAC,WAAW,CAAC,SAAS,GAAG,CAAC,SAAS,CAAC,eAAe;gBAC1E,OAAO,EAAE,GAAG,IAAI,CAAC,OAAO,eAAe,SAAS,EAAE;aACnD;SACF,EACD,SAAS,EACT,KAAK,CACN,CAAA;IACH,CAAC;IAED,OAAO,kBAAkB,CAAC,OAAO,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,EAAE,SAAS,EAAE,KAAK,CAAC,CAAA;AACpE,CAAC;AAED,mEAAmE;AACnE,MAAM,GAAG,GAAG,CAAC,KAAa,EAAU,EAAE,CAAC,CAAC,KAAK,GAAG,IAAI,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAA;AAErE;;;;GAIG;AACH,MAAM,UAAU,eAAe,CAC7B,KAAsB,EACtB,UAA6B,EAC7B,IAAyB,EACzB,UAAsB,EACtB,MAA0B;IAE1B,MAAM,KAAK,GAAG,oBAAoB,CAAC,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,CAAC,CAAA;IAC/E,IAAI,CAAC,KAAK,CAAC,EAAE;QAAE,OAAO,KAAK,CAAC,OAAO,CAAA;IACnC,OAAO,qBAAqB,CAAC,KAAK,CAAC,QAAQ,EAAE,UAAU,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,CAAC,CAAA;AACpF,CAAC"}

package/dist/runtime/managed-models/delete.d.ts

@@ -0,0 +1,15 @@
+import type { ManagedScopePaths } from '../../config/index.js';
+import type { ModelFileOps } from './guest-files.js';
+export interface DeletedManagedModelFiles {
+    freedBytes: number;
+    engineCachesRemoved: number;
+}
+/**
+ * `files` (change `add-tensorrt-llm-windows`, task 2.8): on Windows the model and its caches are in the
+ * WSL guest, and are sized and removed by commands there (`guestModelFiles`) — all caches and the model
+ * in one call each, the caches listed first. Without it, the walk below, as on Linux.
+ */
+export declare function deleteManagedModelFiles(paths: ManagedScopePaths, model: {
+    id: string;
+    dir: string;
+}, files?: ModelFileOps): Promise<DeletedManagedModelFiles>;