Class: GgufInsights

Defined in: gguf/insights/GgufInsights.ts:16

Accessors

ggufFileInfo

Get Signature

get ggufFileInfo(): GgufFileInfo;

Defined in: gguf/insights/GgufInsights.ts:59

Returns

GgufFileInfo

---

configurationResolver

Get Signature

get configurationResolver(): GgufInsightsConfigurationResolver;

Defined in: gguf/insights/GgufInsights.ts:63

Returns

GgufInsightsConfigurationResolver

---

tokens

Get Signature

get tokens(): GgufInsightsTokens;

Defined in: gguf/insights/GgufInsights.ts:67

Returns

GgufInsightsTokens

---

trainContextSize

Get Signature

get trainContextSize(): undefined | number;

Defined in: gguf/insights/GgufInsights.ts:72

The context size the model was trained on

Returns

undefined | number

---

embeddingVectorSize

Get Signature

get embeddingVectorSize(): undefined | number;

Defined in: gguf/insights/GgufInsights.ts:77

The size of an embedding vector the model can produce

Returns

undefined | number

---

totalLayers

Get Signature

get totalLayers(): number;

Defined in: gguf/insights/GgufInsights.ts:81

Returns

number

---

modelSize

Get Signature

get modelSize(): number;

Defined in: gguf/insights/GgufInsights.ts:86

Returns

number

---

flashAttentionSupported

Get Signature

get flashAttentionSupported(): boolean;

Defined in: gguf/insights/GgufInsights.ts:90

Returns

boolean

---

hasEncoder

Get Signature

get hasEncoder(): boolean;

Defined in: gguf/insights/GgufInsights.ts:110

Returns

boolean

---

hasDecoder

Get Signature

get hasDecoder(): boolean;

Defined in: gguf/insights/GgufInsights.ts:120

Returns

boolean

---

isRecurrent

Get Signature

get isRecurrent(): boolean;

Defined in: gguf/insights/GgufInsights.ts:129

Returns

boolean

---

supportsRanking

Get Signature

get supportsRanking(): boolean;

Defined in: gguf/insights/GgufInsights.ts:143

Returns

boolean

---

swaSize

Get Signature

get swaSize(): undefined | number;

Defined in: gguf/insights/GgufInsights.ts:171

The size of the SWA (Sliding Window Attention).

When undefined, the model does not use sliding window attention.

Returns

undefined | number

Methods

getWarnings()

getWarnings(modelFilePath?: string): string[];

Defined in: gguf/insights/GgufInsights.ts:39

Get warnings about the model file that would affect its usage.

Most of these warnings are also generated by llama.cpp

Parameters

Parameter	Type
modelFilePath?	string

Returns

string[]

---

estimateModelResourceRequirements()

estimateModelResourceRequirements(__namedParameters: {
  gpuLayers: number;
  useMmap?: boolean;
  gpuSupportsMmap?: boolean;
}): GgufInsightsResourceRequirements;

Defined in: gguf/insights/GgufInsights.ts:183

Parameters

Parameter	Type
__namedParameters	{ gpuLayers: number; useMmap?: boolean; gpuSupportsMmap?: boolean; }
__namedParameters.gpuLayers	number
__namedParameters.useMmap?	boolean
__namedParameters.gpuSupportsMmap?	boolean

Returns

GgufInsightsResourceRequirements

---

estimateContextResourceRequirements()

estimateContextResourceRequirements(__namedParameters: {
  contextSize: number;
  modelGpuLayers: number;
  batchSize?: number;
  sequences?: number;
  isEmbeddingContext?: boolean;
  flashAttention?: boolean;
  includeGraphOverhead?: boolean;
  swaFullCache?: boolean;
}): GgufInsightsResourceRequirements;

Defined in: gguf/insights/GgufInsights.ts:201

Estimates the memory required to create a context of the given parameters based on the implementation details of llama.cpp. The calculation doesn't include a precise estimation of the graph overhead memory, so it uses a rough estimate for that. The estimation for the graph overhead memory will be improved in the future to be more precise, but it's good enough for now.

Parameters

Parameter	Type
__namedParameters	{ contextSize: number; modelGpuLayers: number; batchSize?: number; sequences?: number; isEmbeddingContext?: boolean; flashAttention?: boolean; includeGraphOverhead?: boolean; swaFullCache?: boolean; }
__namedParameters.contextSize	number
__namedParameters.modelGpuLayers	number
__namedParameters.batchSize?	number
__namedParameters.sequences?	number
__namedParameters.isEmbeddingContext?	boolean
__namedParameters.flashAttention?	boolean
__namedParameters.includeGraphOverhead?	boolean
__namedParameters.swaFullCache?	boolean

Returns

GgufInsightsResourceRequirements

---

from()

static from(ggufFileInfo: GgufFileInfo, llama?: Llama): Promise<GgufInsights>;

Defined in: gguf/insights/GgufInsights.ts:584

Parameters

Parameter	Type	Description
ggufFileInfo	GgufFileInfo
llama?	Llama	If you already have a Llama instance, pass it to reuse it for the GgufInsights instance. If you don't pass a Llama instance, a basic Llama instance is created as a fallback - it's a slim instance that doesn't instantiate a llama.cpp backend, so it won't utilize the GPU at all, and be shared with other GgufInsights instances that need a fallback Llama instance.

Returns

Promise<GgufInsights>